International Conference on Artificial Intelligence and Sustainable Engineering: Select Proceedings of AISE 2020, Volume 1 (Lecture Notes in Electrical Engineering, 836) 981168541X, 9789811685415
This book comprises select papers from the International Conference on Artificial Intelligence and Sustainable Engineeri
140 100 15MB
English Pages 581 [555] Year 2022
Preface
Contents
About the Editors
Smart Grid and Net Metering Impact in Current Scenario of Power System
1 Introduction
2 Smart Grid—Future of Smart City
2.1 Substation Automation
2.2 Advanced Meter Reading
2.3 Advanced Meter Infrastructure
2.4 Distribution Automation
2.5 Demand Response
2.6 Energy Management System
3 Smart Meters
3.1 Advantages of Smart Meters
3.2 Methodology Adopted by EESL
3.3 Components Inside a Smart Meter
4 Grid Stabilization Using PMU—“Phasor Measurement Unit”
5 Net Metering
5.1 Types of Net Metering
6 Conclusion
7 Future Scope
References
Machine Learning-Based Approach for Myocardial Infarction
1 Introduction
2 Materials and Methods
3 Results and Discussion
4 Conclusion and Future Scope
References
Joyful Jaccard: An Analysis of Jaccard-Based Similarity Measures in Collaborative Recommendations
1 Introduction
2 Related Work
3 Jaccard-Based Similarity Measures
3.1 Jmsd
3.2 Cjmsd
3.3 Relevant Jaccard (RJaccard)
3.4 JacRA
3.5 JacLMH
3.6 CJacMD
3.7 New Weighted Similarity Measure (NWSM)
3.8 New Heuristic Similarity Model (NHSM)
4 Experiments
4.1 Evaluation Metrics
4.2 Results and Discussion
5 Conclusion
References
An Improved Reversible Data Hiding Technique For Encrypted Images Using Difference Error Expansion
1 Introduction
2 Prediction Error Expansion
3 Proposed Scheme
3.1 Encryption Phase
3.2 Data Embedding Phase
3.3 Extraction Phase
4 Experimental Analysis
4.1 Based on Payload Limit
4.2 Based on Security
4.3 Comparison
5 Conclusion
References
Assessment of Image Quality Metrics by Means of Various Preprocessing Filters for Lung CT Scan Images
1 Introduction
2 Methodology
2.1 Input CT Scan Image
2.2 Filtering Techniques
2.3 Performance Parameter Measure
2.4 Results and Discussion
2.5 Conclusion
References
A New Approach for Acute Lymphocytic Leukemia Identification Using Transfer Learning
1 Introduction
2 Previous Work
3 Dataset and Background
3.1 The Dataset
3.2 Classification Models
3.3 SVM Classifier
3.4 Random Forest Classifier
4 Proposed Methodology
5 Results
6 Conclusions and Future Scope
References
A Hybrid Machine Learning Approach for Customer Segmentation Using RFM Analysis
1 Introduction
2 Business Background and the Associated Data
3 Data Preprocessing
4 RFM Model-based Clustering Analysis
5 Enhancing K-Means Clustering Analysis Using Decision Tree
6 Customer Centric Business Intelligence and Recommendations
7 Concluding Remarks
References
AMD-Net: Automatic Medical Diagnoses Using Retinal OCT Images
1 Introduction
2 Related Work
3 Dataset
4 Proposed Work
5 Experiment Result and Discussion
6 Conclusion
References
Performance Interpretation of Supervised Artificial Neural Network Highlighting Role of Weight and Bias for Link Prediction
1 Introduction
2 Related Work
3 Data Sources and Evaluation Methods
4 Experimental Work and Results
5 Conclusion
References
Using Big Data Analytics on Social Media to Analyze Tourism Service Encounters
1 Introduction
2 Literature Review
3 Research Methodology
3.1 Data Collection
3.2 Data Preprocessing
3.3 Data Analysis Methodology
4 Results and Discussion
4.1 Insights from Sentiment Analysis
4.2 Insights from Topic Modeling
5 Conclusion
References
Measuring the Efficiency of LPWAN in Disaster Logistics System
1 Introduction
2 Related Work
2.1 Disaster Logistics Assistance System (DLAS)
2.2 Communication for the Logistics System
3 Proposed Scheme
3.1 LoRa Technology
4 Analysis and Simulation
4.1 Comparison Between NB-IoT and LoRa
4.2 Latency
4.3 Geolocation and Density
4.4 Power Usage and QoS
4.5 Supply Chain Tracking
4.6 Testing Ground
5 Conclusion
References
Automatic Classroom Monitoring System Using Facial Expression Recognition
1 Introduction
2 Literature Review
3 Methodology
3.1 Data Capturing
3.2 Face Detection
3.3 Face Recognition
3.4 Emotion Recognition
3.5 Emotion Analysis of Classroom
4 Experimental Setup
4.1 Dataset
4.2 Model Implementation
5 Discussion
References
Performance Evaluation of Image-Based Diseased Leaf Identification Model Using CNN and GA
1 Introduction
2 Dataset Description
3 Algorithm Used
3.1 Convolution NN
3.2 Genetic Algorithm
3.3 Next-Generation
4 Result
4.1 Result Analysis
5 Conclusion
References
Urinary System Diseases Prediction Using Supervised Machine Learning-Based Model: XGBoost and Random Forest
1 Introduction
2 Literature Survey
3 Research Methodology
3.1 About Dataset and Preprocessing
3.2 System Model
4 Result
5 Conclusion
References
Software Product Line Testing—A Proposal of Distance-Based Approach
1 Introduction
2 Background
3 Related Work
4 Distance-Based Testing Method for SPL
5 Experiment and Result Evaluation
6 Validity Threats
7 Conclusion and Future Work
References
ANFIS-Based STATCOM for Reactive Power Compensation of Dynamic Loads Under Microgrid Disturbances
1 Introduction
2 System Description
3 Controllers for STATCOM
4 Results and Discussions
5 Conclusion
References
Evaluation of Series Voltage Injection Using Genetic Algorithm for Constant Load End Voltage
1 Introduction
2 Series Voltage Injection Planning
2.1 Nominal T-model [Case 1: Vinj at Supply End]
2.2 Nominal T-model [Case 2: Vinj at Load End]
2.3 Nominal Pi-Model [Case 1: Vinj at Supply End]
2.4 Nominal Pi-Model [Case 2: Vinj at Load End]
3 Genetic Algorithm
4 Results and Discussion
5 Conclusion
References
Human Bone Assessment: A Deep Convolutional Neural Network Approach
1 Introduction
2 Related Works
3 the Proposed Method
4 Result
5 Discussion
6 Conclusion
References
Review on Pulse Width Modulation and Optimization Techniques for Multilevel Inverters
1 Introduction
2 Multilevel Inverters
3 Multilevel Inverters Topologies
3.1 Cascaded H-Bridge Multilevel Inverters.
3.2 Diode-Clamped Multilevel Inverters
3.3 Capacitor-Clamped Multilevel Inverters
4 Modulation Techniques
4.1 Sinusoidal Pulse Width Modulation (SPWM)
4.2 Multilevel Space Vector Modulation
4.3 Selective Harmonic Elimination PWM
5 Optimization Techniques
5.1 Angle Control Technique Optimization
5.2 Minimization Using Optimization
5.3 Genetic Algorithm Technique
6 Comparison of Various Techniques
7 Particle Swarm Optimization
8 Conclusion
References
Nature-Inspired AI Techniques in Intelligent Transportation System
1 Introduction
2 Nature-Inspired AI Techniques
3 AI-Based ITS Applications
3.1 Road Freight Transportation
3.2 Route and Network Planning for Traffic Management
3.3 Vehicle Control System (Smart Vehicle)
3.4 Transportation Safety and Security
4 Future Research Directions
5 Conclusions
References
Determining Optimal Epsilon (eps) on DBSCAN Using Empty Circles
1 Introduction
2 Preliminaries
2.1 Convex Hull
2.2 Voronoi Diagram
2.3 Voronoi Circle/Empty Circle
2.4 DBSCAN Algorithm
2.5 K-Nearest Neighbor’s Algorithm
2.6 Euclidean Distance and Completeness
3 The Proposed Algorithm
4 Experimental Results
5 Discussions
6 Conclusions
References
A Hybrid Design for Low-Power Fault Tolerant One-Bit Full Adder for Neural Network Applications
1 Introduction
2 Literature Survey
2.1 Fault Tolerant Techniques
2.2 Low-Power Techniques
3 Proposed Design
3.1 Self-checking Full Adder
3.2 Self-repairing Full Adder
4 Implementation
4.1 16T One-Bit Full Adder
4.2 6T-XOR Gate
4.3 2 × 1 Multiplexer
4.4 Self-checking Adder
4.5 Self-repairing Full Adder
5 Results and Discussion
5.1 Simulation Results
5.2 Comparisons
6 Conclusion
References
Modified ResNet for Volumetric Segmentation of Liver and Its Tumor from 3D CT
1 Introduction
2 Related Work
3 Methodology
3.1 Materials
3.2 Preprocessing
3.3 Modified ResNet
4 Results and Discussion
5 Conclusion
References
A Study on Effects of Different Image Enhancement Techniques on Cervical Colposcopy Images
1 Introduction
2 Background
3 Methods and Material
3.1 Data Collection
3.2 Overview of the Work
4 Evaluation Metrics
5 Results
6 Discussion
7 Conclusion
References
Fully Decentralized Blockchain and Browser-Based Volunteer Computing Platform
1 Introduction
2 Related Work
3 Motivation and Background
4 Existing Model
5 Design Goals
6 Proposed Architecture
6.1 Task Submitter and Executor
6.2 Blockchain Incentivization Mechanism
6.3 Smart Contract Task Scheduling
6.4 Browser-Based Computing
7 Flow of Control
8 Results
9 Analysis
10 Conclusion
References
Generation of Nepalese Handwritten Characters Using Generative Adversarial Network
1 Introduction
2 Research Methodology
2.1 Dataset Collection and Preparation
2.2 Designing Deep Convolutional Generative Adversarial Networks
2.3 Validation
2.4 Tools and IDE
2.5 Test Bed Configuration
3 Result and Analysis
3.1 Experiment with Optimizer: Adam
3.2 Experiment with Optimizer: RMSProp
3.3 Comparison of Results
4 Conclusion
References
An Intelligent Code Smell Detection Technique Using Optimized Rule-Based Architecture for Object-Oriented Programmings
1 Introduction
1.1 Bad Smell Types and Detection Tools
1.2 Detection Tools
1.3 Machine Learning
1.4 Multi-Label Classification Process
1.5 ALO (Ant Lion Optimization)
2 Related Work
3 Methodology
4 Results and Discussion
5 Conclusion and Future Scope
References
An Evaluation of LDA Modeling in Financial News Articles
1 Introduction
2 Data Selection and Preprocessing
3 Variation of Latent Dirichlet Allocation (LDA)
4 Evaluation Metrics
4.1 Perplexity
4.2 Coherence
4.3 Word2Vec-Based Similarity Score
5 Result
6 Future Work and Conclusion
References
Analysis of Online Toxicity Detection Using Machine Learning Approaches
1 Introduction
1.1 What is OHS and Origin of It?
1.2 Perpetrator Mission and Consequences
2 Related Work
3 Data and Evaluation
3.1 Dataset
3.2 Preprocessing
3.3 Feature Selection
3.4 Models
4 Results
4.1 Experiment 1: Using tf-idf Only
4.2 Experiment 2: Using tf-idf on POS Tags
4.3 Experiment 3: Using Sentence Embeddings
5 Conclusion
References
Highly Precise ANN Classifier for Pancreatic Tumor Recognition with Fuzzy C-means Segmentation
1 Introduction
2 Recent Works
3 Proposed Works
3.1 Median Filter
3.2 Fuzzy C-mean Segmentation
3.3 Feature Extraction Using SIFT
3.4 Threshold Method
3.5 Artificial Neural Network Classifier
4 Result and Discussion
5 Conclusion
References
GA-ANN Framework for Breast Cancer Classification Using NSGA-II
1 Introduction
2 Literature Review
3 Material and Methods
3.1 Non-dominated Sorting Genetic Algorithm (NSGA-II)
3.2 Formulation of the Proposed Approach Using Multi-objective Optimization
3.3 Mathematical Model of Proposed Approach
4 Results and Analysis
4.1 Experimental Results
5 Conclusion
References
Applying Support Vector Machine Algorithm in Diabetes Prediction in Indian Context
1 Introduction
2 Literature Survey
3 Proposed Methodology and Results
4 Conclusion and Future Scope
References
AI Technologies and Firefly Algorithms
1 Introduction
1.1 Bio-intelligence-inspired Algorithms
1.2 Artificial Intelligence
1.3 Organization of the Paper
2 Related Surveys
3 Firefly Algorithm
4 Latest Developments in Firefly Algorithms
5 Artificial Intelligence in Firefly Algorithm
5.1 Artificial Neural Networks
5.2 Computer Vision
5.3 Fuzzy Logic
5.4 Nature Language Processing
5.5 Support Vector Machine
6 Future Directions
7 Conclusion
References
Detection and Classification of Fetal Heart Rate (FHR)
1 Introduction
2 Related Work
3 Proposed System
4 Methodology
4.1 Dataset
4.2 Signal Analysis
4.3 Extracting Filter Parameters
4.4 Shannon Energy Envelope-Based Beat Localization
4.5 Final Dataset Creation and Preprocessing
4.6 Choosing a Classifier
4.7 Fetal Health Classification
5 Results
6 Conclusion
7 Future Scope
References
Performance Analysis of Fog Computing Through Non-Markovian Queuing System
1 Introduction
2 Related Work
3 System Model
3.1 Assumptions
3.2 Queueing Model
4 Simulation Results
5 Discussion
6 Conclusion
References
Cancer Classification Based on an Integrated Clustering and Classification Model Using Gene Expression Data
1 Introduction
2 Related Works
3 Overview of the Approach
3.1 Preprocessing
3.2 Feature Selection
3.3 Clustering
3.4 Reduction of the Feature Subset
3.5 Classification
4 Result and Discussion
4.1 Description of Experimental Data
4.2 Comparison of the Proposed Integrated Technique with Other Models
4.3 Comparison Between the Proposed Model and the One Without the Clustering Approach
5 Conclusion
References
A Vocational Career Advisory Application Built Using Unsupervised Machine Learning Frameworks
1 Introduction
2 Process Flow of the System
2.1 Dataset Description
2.2 Making the Data Usable
2.3 Encoding the Data Label
3 Machine Learning Algorithm Used
3.1 Support Vector Machines
3.2 Decision Trees
4 Training and Testing
4.1 Prediction Using Trained Model
4.2 Creating a User-Friendly GUI
4.3 Display Current Job Opening
5 Result
6 Future Scope
References
UML-Based Modelling for Legal Rule Using Natural Language Processing
1 Introduction
2 Fundamentals of Natural Language Processing (NLP)
3 Fundamentals of Unified Modelling Language (UML)
4 Proposed Model
5 UML-Based Modelling of Proposed Legal System
5.1 Class Diagram
5.2 Program Snippet
5.3 Association Diagram
5.4 Sample Output
6 Future Direction
7 Conclusion
References
Deep Learning Algorithm for Procedure and Network Inference for Genomic Data
1 Introduction
2 Background
3 Datasets and Validation Procedure
4 Results Obtained
5 Conclusions
References
Encryption and Decryption of Secure Data for Diverse Genomes
1 Introduction
2 Literature Review
3 Implementation of Proposed Algorithm
4 Results and Discussions
5 Conclusion
References
Survey on Sentiment Analysis for Mix Code Text
1 Introduction
2 Literature Survey
3 Datasets
3.1 Build Own Dataset
3.2 Evaluation Datasets
4 Conclusion
References
Comparison of Visual Question Answering Datasets for Improving Their Reading Capabilities
1 Introduction
2 TextVQA
3 ST-VQA
4 OCR-VQA-200 K
5 Conclusion
References
Underwater Acoustic Image Processing for Detection of Marine Debris
1 Introduction
2 Underwater Image Acquisition
2.1 Acoustic Imaging
2.2 Acoustic Image Formation
2.3 Characteristics of Acoustic Images
3 Image Enhancement and Restoration Techniques
3.1 Spatial Domain Methods
3.2 Transformation Domain
3.3 Filtering Techniques
4 Experiment and Result
4.1 Dataset
4.2 Methodology
4.3 Performance Measures
5 Conclusion
References
An IoT Approach Toward Storage of Medicines to Develop a Smart Pill Box
1 Introduction
2 Literature Survey
3 Problem Statement
4 Smart Medical Box
4.1 Methodology
4.2 The Device
4.3 The Android Application
4.4 Program Code
5 Experimental Results
6 Conclusion and Future Scope
References
Artificial Neural Network-Based Model for the Prediction of Evaporation in Agriculture
1 Introduction
2 Materials and Methods
2.1 Neural Network Structure Used in the Literature
2.2 Proposed System
2.3 Activation Function
2.4 Optimizers
2.5 Design Flow For Evaporation Prediction
3 Implementation
4 Results
5 Conclusion
6 Future Scope
References
Recommend Papers
File loading please wait...
Citation preview
Lecture Notes in Electrical Engineering 836
Goutam Sanyal · Carlos M. Travieso-González · Shashank Awasthi · Carla M. A. Pinto · B. R. Purushothama Editors
International Conference on Artificial Intelligence and Sustainable Engineering Select Proceedings of AISE 2020, Volume 1
Lecture Notes in Electrical Engineering Volume 836
Series Editors Leopoldo Angrisani, Department of Electrical and Information Technologies Engineering, University of Napoli Federico II, Naples, Italy Marco Arteaga, Departament de Control y Robótica, Universidad Nacional Autónoma de México, Coyoacán, Mexico Bijaya Ketan Panigrahi, Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, Delhi, India Samarjit Chakraborty, Fakultät für Elektrotechnik und Informationstechnik, TU München, Munich, Germany Jiming Chen, Zhejiang University, Hangzhou, Zhejiang, China Shanben Chen, Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China Tan Kay Chen, Department of Electrical and Computer Engineering, National University of Singapore, Singapore, Singapore Rüdiger Dillmann, Humanoids and Intelligent Systems Laboratory, Karlsruhe Institute for Technology, Karlsruhe, Germany Haibin Duan, Beijing University of Aeronautics and Astronautics, Beijing, China Gianluigi Ferrari, Università di Parma, Parma, Italy Manuel Ferre, Centre for Automation and Robotics CAR (UPM-CSIC), Universidad Politécnica de Madrid, Madrid, Spain Sandra Hirche, Department of Electrical Engineering and Information Science, Technische Universität München, Munich, Germany Faryar Jabbari, Department of Mechanical and Aerospace Engineering, University of California, Irvine, CA, USA Limin Jia, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Alaa Khamis, German University in Egypt El Tagamoa El Khames, New Cairo City, Egypt Torsten Kroeger, Stanford University, Stanford, CA, USA Yong Li, Hunan University, Changsha, Hunan, China Qilian Liang, Department of Electrical Engineering, University of Texas at Arlington, Arlington, TX, USA Ferran Martín, Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, Bellaterra, Barcelona, Spain Tan Cher Ming, College of Engineering, Nanyang Technological University, Singapore, Singapore Wolfgang Minker, Institute of Information Technology, University of Ulm, Ulm, Germany Pradeep Misra, Department of Electrical Engineering, Wright State University, Dayton, OH, USA Sebastian Möller, Quality and Usability Laboratory, TU Berlin, Berlin, Germany Subhas Mukhopadhyay, School of Engineering & Advanced Technology, Massey University, Palmerston North, Manawatu-Wanganui, New Zealand Cun-Zheng Ning, Electrical Engineering, Arizona State University, Tempe, AZ, USA Toyoaki Nishida, Graduate School of Informatics, Kyoto University, Kyoto, Japan Federica Pascucci, Dipartimento di Ingegneria, Università degli Studi “Roma Tre”, Rome, Italy Yong Qin, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China Gan Woon Seng, School of Electrical & Electronic Engineering, Nanyang Technological University, Singapore, Singapore Joachim Speidel, Institute of Telecommunications, Universität Stuttgart, Stuttgart, Germany Germano Veiga, Campus da FEUP, INESC Porto, Porto, Portugal Haitao Wu, Academy of Opto-electronics, Chinese Academy of Sciences, Beijing, China Walter Zamboni, DIEM - Università degli studi di Salerno, Fisciano, Salerno, Italy Junjie James Zhang, Charlotte, NC, USA
The book series Lecture Notes in Electrical Engineering (LNEE) publishes the latest developments in Electrical Engineering - quickly, informally and in high quality. While original research reported in proceedings and monographs has traditionally formed the core of LNEE, we also encourage authors to submit books devoted to supporting student education and professional training in the various fields and applications areas of electrical engineering. The series cover classical and emerging topics concerning: • • • • • • • • • • • •
Communication Engineering, Information Theory and Networks Electronics Engineering and Microelectronics Signal, Image and Speech Processing Wireless and Mobile Communication Circuits and Systems Energy Systems, Power Electronics and Electrical Machines Electro-optical Engineering Instrumentation Engineering Avionics Engineering Control Systems Internet-of-Things and Cybersecurity Biomedical Devices, MEMS and NEMS
For general information about this book series, comments or suggestions, please contact [email protected]. To submit a proposal or request further information, please contact the Publishing Editor in your country: China Jasmine Dou, Editor ([email protected]) India, Japan, Rest of Asia Swati Meherishi, Editorial Director ([email protected]) Southeast Asia, Australia, New Zealand Ramesh Nath Premnath, Editor ([email protected]) USA, Canada: Michael Luby, Senior Editor ([email protected]) All other Countries: Leontina Di Cecco, Senior Editor ([email protected]) ** This series is indexed by EI Compendex and Scopus databases. **
More information about this series at https://link.springer.com/bookseries/7818
Goutam Sanyal · Carlos M. Travieso-González · Shashank Awasthi · Carla M. A. Pinto · B. R. Purushothama Editors
International Conference on Artificial Intelligence and Sustainable Engineering Select Proceedings of AISE 2020, Volume 1
Editors Goutam Sanyal National Institute of Technology (NIT) Durgapur, West Bengal, India Shashank Awasthi Department of Computer Science and Engineering GL Bajaj Institute of Technology and Management Greater Noida, India
Carlos M. Travieso-González University of Las Palmas de Gran Canaria Las Palmas de Gran Canaria, Spain Carla M. A. Pinto School of Engineering, Polytechnic of Porto University of Porto Porto, Portugal
B. R. Purushothama Department of Planning and Development National Institute of Technology Goa Goa, India
ISSN 1876-1100 ISSN 1876-1119 (electronic) Lecture Notes in Electrical Engineering ISBN 978-981-16-8541-5 ISBN 978-981-16-8542-2 (eBook) https://doi.org/10.1007/978-981-16-8542-2 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore
Preface
Artificial intelligence (AI) has accelerated progress in every sphere of a human’s life. AI is also helping the next generation of companies to reduce their environmental and social impact by improving efficiency and developing new products. But, it also brings greater challenges to the sustainable development of engineering products. Sustainability is the greatest challenge in a variety of areas like AI, AR/VR, robotics, IoT, non-conventional energy and environment, agriculture, health, transportation, etc. Therefore, the relationship between AI and sustainable engineering is worth studying. This book presents select proceedings of the International Conference on Artificial Intelligence and Sustainable Engineering (AISE-2020). It covers various topics like artificial intelligence in security and surveillance, health care, big data analytics, engineering design for sustainable development using IoT/AI, etc. This book can be a valuable resource for academicians, researchers, and professionals working in the field of artificial intelligence and can provide solutions to the challenges faced in sustainable engineering based on AI and supporting tools. It will contribute in enhancing the understanding of knowledge and research related issues in the domain of Artificial Intelligence and Sustainable Engineering. Durgapur, India Las Palmas, Spain Noida, India Porto, Portugal Goa, India
Goutam Sanyal Carlos M. Travieso-González Shashank Awasthi Carla M. A. Pinto B. R. Purushothama
v
Contents
Smart Grid and Net Metering Impact in Current Scenario of Power System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sarah Biswal and Divya Asija Machine Learning-Based Approach for Myocardial Infarction . . . . . . . . . Pooja Maindarkar and S. Sofana Reka
1 17
Joyful Jaccard: An Analysis of Jaccard-Based Similarity Measures in Collaborative Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Anisha Jindal, Naveen Sharma, and Vijay Verma
29
An Improved Reversible Data Hiding Technique For Encrypted Images Using Difference Error Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . Ankita Vaish, Ruchi Agarwal, and Manoj Kumar
43
Assessment of Image Quality Metrics by Means of Various Preprocessing Filters for Lung CT Scan Images . . . . . . . . . . . . . . . . . . . . . . Sugandha Saxena, S. N. Prasad, and T. S. Deepthi Murthy
59
A New Approach for Acute Lymphocytic Leukemia Identification Using Transfer Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Saba Farheen Munshi and Chandrakant P. Navdeti
71
A Hybrid Machine Learning Approach for Customer Segmentation Using RFM Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Poonam Chaudhary, Vaishali Kalra, and Srishti Sharma
87
AMD-Net: Automatic Medical Diagnoses Using Retinal OCT Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 Praveen Mittal Performance Interpretation of Supervised Artificial Neural Network Highlighting Role of Weight and Bias for Link Prediction . . . . . 109 Sandhya Pundhir, Varsha Kumari, and Udayan Ghose
vii
viii
Contents
Using Big Data Analytics on Social Media to Analyze Tourism Service Encounters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Sunil Kumar, Arpan Kumar Kar, and P. Vigneswara Ilavarasan Measuring the Efficiency of LPWAN in Disaster Logistics System . . . . . . 131 Nitin Rastogi and Aravendra Kumar Sharma Automatic Classroom Monitoring System Using Facial Expression Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Sadhana Singh, Aaryan Gupta, and R. S. Pavithr Performance Evaluation of Image-Based Diseased Leaf Identification Model Using CNN and GA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Anita Shrotriya, Akhilesh Kumar Sharma, and Jyoti Grover Urinary System Diseases Prediction Using Supervised Machine Learning-Based Model: XGBoost and Random Forest . . . . . . . . . . . . . . . . 179 Atul Kumar Uttam Software Product Line Testing—A Proposal of Distance-Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 Ashish Saini, Rajkumar, and Satendra Kumar ANFIS-Based STATCOM for Reactive Power Compensation of Dynamic Loads Under Microgrid Disturbances . . . . . . . . . . . . . . . . . . . . 199 Raman Prajapati and Sheela Tiwari Evaluation of Series Voltage Injection Using Genetic Algorithm for Constant Load End Voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 Vishal R. Wagh and Atul S. Koshti Human Bone Assessment: A Deep Convolutional Neural Network Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 D. P. Yadav Review on Pulse Width Modulation and Optimization Techniques for Multilevel Inverters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 S. Vedith, J. N. Bhanu Tej, S. Sampath, M. Usha Sree, P. Nithin Rao, and K. Neelima Nature-Inspired AI Techniques in Intelligent Transportation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 Megha Mahobe, Pradeep Kumar, and Shashi Shekhar Jha Determining Optimal Epsilon (eps) on DBSCAN Using Empty Circles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 Kinsuk Giri and Tuhin Kr. Biswas A Hybrid Design for Low-Power Fault Tolerant One-Bit Full Adder for Neural Network Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277 C. Raji and S. N. Prasad
Contents
ix
Modified ResNet for Volumetric Segmentation of Liver and Its Tumor from 3D CT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295 Gajendra Kumar Mourya, Dinesh Bhatia, Manashjit Gogoi, S. N. Talbar, Ujjwal Baid, and Prasad Dudante A Study on Effects of Different Image Enhancement Techniques on Cervical Colposcopy Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303 Elima Hussain, Lipi B. Mahanta, Khurshid A. Borbora, Ankit Kumar Shah, Divya Subhasini, and Tarali Das Fully Decentralized Blockchain and Browser-Based Volunteer Computing Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 K. S. Sagar Bharadwaj, Samvid Dharanikota, Adarsh Honawad, D. Usha, and K. Chandrasekaran Generation of Nepalese Handwritten Characters Using Generative Adversarial Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333 Krishna Khadka, Sanjeeb Prasad Panday, and Basanta Joshi An Intelligent Code Smell Detection Technique Using Optimized Rule-Based Architecture for Object-Oriented Programmings . . . . . . . . . . 349 Manpreet Kaur and Daljeet Singh An Evaluation of LDA Modeling in Financial News Articles . . . . . . . . . . . 365 Rahul Katarya, Indrajeet Das, Bhanu Shrivastava, and Karan Keswani Analysis of Online Toxicity Detection Using Machine Learning Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381 Anjum and Rahul Katarya Highly Precise ANN Classifier for Pancreatic Tumor Recognition with Fuzzy C-means Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 Ajanthaa Lakkshmanan and C. Anbu Ananth GA-ANN Framework for Breast Cancer Classification Using NSGA-II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407 Mallikarjuna Rao Gundavarapu, M. Divya Satya Padma, Ch. Mallikarjuna Rao, D. V. Lalitha Parameswari, and G. Saaketh Koundinya Applying Support Vector Machine Algorithm in Diabetes Prediction in Indian Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419 Narendra Mohan and Vinod Jain AI Technologies and Firefly Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425 A. Albert Raj, S. Ravi, and M. Joseph Detection and Classification of Fetal Heart Rate (FHR) . . . . . . . . . . . . . . . 437 Emad Haque, Tanishka Gupta, Vinayak Singh, Kaustubh Nene, and Akhil Masurkar
x
Contents
Performance Analysis of Fog Computing Through Non-Markovian Queuing System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449 K. Gokulnath Cancer Classification Based on an Integrated Clustering and Classification Model Using Gene Expression Data . . . . . . . . . . . . . . . . 461 Ananya Das and Subhashis Chatterjee A Vocational Career Advisory Application Built Using Unsupervised Machine Learning Frameworks . . . . . . . . . . . . . . . . . . . . . . . 471 Ashutosh Shankhdhar, Anurag Gupta, Akhilesh Kumar Singh, and Rahul Pradhan UML-Based Modelling for Legal Rule Using Natural Language Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481 Tanaya Das, Riya Sil, Abhishek Roy, and Arun Kumar Majumdar Deep Learning Algorithm for Procedure and Network Inference for Genomic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493 Raveendra Gudodagi, R. Venkata Siva Reddy, and Mohammed Riyaz Ahmed Encryption and Decryption of Secure Data for Diverse Genomes . . . . . . . 505 Raveendra Gudodagi and R. Venkata Siva Reddy Survey on Sentiment Analysis for Mix Code Text . . . . . . . . . . . . . . . . . . . . . 515 Rahul Pradhan and Dilip Kumar Sharma Comparison of Visual Question Answering Datasets for Improving Their Reading Capabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525 Himanshu Sharma and Anand Singh Jalal Underwater Acoustic Image Processing for Detection of Marine Debris . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535 Vritika Vijaylal Naik and Sadaf Ansari An IoT Approach Toward Storage of Medicines to Develop a Smart Pill Box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547 Amara Aditya Manikanta, Himanshu Sahu, Kratika Arora, and SriKrishna Vamsi Koneru Artificial Neural Network-Based Model for the Prediction of Evaporation in Agriculture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561 Kamal Batra and Parul Gandhi
About the Editors
Dr. Goutam Sanyal has served as the Head of the Department of Computer Science & Engineering at the National Institute of Technology (NIT) Durgapur, India. He received B.Tech. and M.Tech. from NIT, Durgapur, and Ph.D. (Engg.) from Jadavpur University, Kolkata, in robot manipulator path planning. He holds over 35 years of experience in the field of teaching, research, and administration. He has published nearly 200 papers in reputed international journals and conferences. He has guided 22 Ph.D. scholars in steganography, wireless sensor networks, computer vision, natural language processing. He has supervised more than 40 PG and 300 UG theses. He is a co-author of a book in Computer Graphics and Multimedia and 8 Book chapters. He is a regular Member of IEEE, a Life Member of CSI, and a Fellow of IEI. His biography has been selected for inclusion in Marquis Who’s Who in the World 2016, 2017, 2018, 2019, and 2020 Edition. Carlos M. Travieso-González is a Full Professor of signal processing and pattern recognition and head of the signals and communications department at the University of Las Palmas de Gran Canaria (ULPGC-Spain). He received the M.Sc. degree in 1997 in Telecommunication Engineering at Polytechnic University of Catalonia (UPC), Spain, and a Ph.D. degree in 2002 at ULPGC. His research lines are biometrics, biomedical signals and images, data mining, classification system, signal and image processing, machine learning, and environmental intelligence. He has researched 51 international and Spanish research projects. He has 04 authored books, 24 edited books, 440 journal papers, and 07 patents in Spanish Patent and Trademark Office published to his credit. He has been a supervisor on 8 Ph.D. Thesis (12 more are under supervision), and 130 Master Thesis. Dr. Shashank Awasthi is a professor in the computer science and engineering department of GL Bajaj Institute of Technology & Management, India. He holds a Ph.D. degree in Computer Science & Engineering and M.Tech. in Computer Science & Engineering from Dr. APJ Kalam Technical University, Lucknow, and MCA from Dr. BR Ambedkar University Agra. His area of interest is wireless sensor networks. He is having more than 18 years of experience in teaching and research. He has xi
xii
About the Editors
published over 30 research papers in international journals/conferences of repute. He is a member of IEEE and the International Association of Engineers, Hong Kong. He is a member of the Editorial Board of various reputed International Journals. Carla M. A. Pinto is an adjunct professor at the school of engineering at Polytechnic of Porto. She completed her Doctorate in Mathematics in 2004 from Universidade do Porto Faculdade de Ciências, Portugal. Prof. Pinto published over 46 articles in highimpact peer-review journals and more than 50 in international peer-reviewed conferences. She works in the area of applied mathematics with a focus on epidemiology and robotics. B. R. Purushothama obtained his Ph.D. in computer science and engineering from the National Institute of Technology (NIT) Warangal, India, and his M.Tech. in computer science and engineering from the NIT Surathkal, India. He is currently working as an assistant professor of computer science and engineering at NIT Goa, India. He has academic experience of over 16 years. His areas of interest are cryptography & information security, security analytics, cloud security, machine learning applications to Security. He has published several works in peer-reviewed journals and proceedings.
Smart Grid and Net Metering Impact in Current Scenario of Power System Sarah Biswal and Divya Asija
Abstract The conventional grid has more Aggregate Technology and Commercial losses (AT&C) but with the evolution of smart grids a drastic change came to the system. This introduces with a two-way communication where not only information and data are exchanged between the utility and its consumers but also electricity exchange also takes place. It is a developing network of communications that controls computers automation and new technologies and tools working together so as to make the grid more efficient, more reliable, more secure and producing greener energy. The smart grid enables newer technologies to be integrated such as wind, thermal and solar power production. From adopting to new changes in our power system to the challenge of grid stabilization using PMU. With the idea of using the intermittent sources, they discovered a technique which not only enhances the power generation technique but also enables the user to generate and consume the same energy. The user can generate energy using PV solar panel and consume that generated power, not only the user consume that generated power but also can give the unconsumed power back to the grid. This helps the utility to manage the power/energy demand in those places where requirement of energy is more. In this way, there is no power outages, and no harm to environment is caused as the energy that is generated is “green energy.” This paper presents the evolution of smart grid and net metering system taken by Indian Ministry of Power to implement Smart Grid in Indian Power System. Keywords AT & C · Smart grid · Utility · PMU · Net metering
1 Introduction India is the second largest populated country in the world; the total population recorded was 132.26 crores (according to 2018 census) and seventh largest populous country. For such countries where population is very large, the power consumption S. Biswal · D. Asija (B) Amity University, Noida, UP, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_1
1
2
S. Biswal and D. Asija
and demand will be more. As of 2018 reports, India among all the other countries in the world was recorded as the third largest consumer of electricity [1]. Over 23 million houses in rural places in India are without electricity. They still do not have the access to electricity and live in complete darkness. Some of the prominent problems which eventually lead to the situation of blackout are: i. ii. iii.
Aging infrastructure of our power grids Losses due to Aggregate Technology and Commercial (AT&C) Revenue generated by Distribution Companies (DISCOMs) is not enough.
These all problems were faced due to the conventional power grid. If we want to solve the existing issues, we should make drastic changes to the system [2]. Thus, the requirement of smart grid has become necessary in our country. A twoway communication is established between the utility and its consumers; the sensing along the transmission line makes grid smart also boosts the power efficiency and lowers the expenditure. It has an infrastructure which can network one or more parts of smart grid via secured high-speed bandwidth connections [3]. It is supposed to implement the smart grid from the ground up, starting with Low Voltage (LV) substations, smart meters and streetlights. As soon as we complete these additives in the grid, we can look upon the leakages, offer streetlight dimming, smart automation, load balancing and many other capabilities that smart grid offers. Energy is the prime input in economic activities and developments in the country. There is still time left to transform our energy consumption and production, what we consume needs to be done wisely and carefully, how we produce energy, what methods we use to generate energy keeping in mind that in future there should not be any scarcity of energy sources. Smart grid allows the consumer to use and generate energy, green and cleaner energy [4] using solar panel setting up at their homes and using that energy in their household and giving up the excess energy back to the grid [5]. Figure 1 shows the analysis for ten years of electricity consumption and demand by the consumers in India which is linearly increasing for each financial year. The main reason for the increase is population and economic development. Analyzing the graph, it is estimated that there will be 90% of energy demand or growth till the year 2035.
2 Smart Grid—Future of Smart City When we talk about smart city, it is defined as one that has an effective project or a plan in eight functional areas of the city. Every parameter has specific module that define the smartness of the city. One of the key parameters for the contribution toward the smart city is “energy sector” [7]. Smart grid and related developments are the important aspects of smart city [8]. It allows the power grid companies to manage the complete blackout (outage) by using various technologies, data analytics. Further, depending on the imminent renewable sources like solar rooftops can be a
Smart Grid and Net Metering Impact in Current …
3
UNTS GENERATED
Electricity Generated (Conventional Sources) Year Wise 1400 1200 1000 800 600 400 200 0
2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 Series1 771.55811.14876.89912.15967.151048.71112.21160.11166.11249.3 ELECTRICITY GENERATED
Fig. 1 Survey of electricity generated over past ten years [6]
game changer in the system of energy sector, further we discuss about the usage of imminent renewable sources in the section “net metering” [4]. The different parts of the network are as discussed below [9].
2.1 Substation Automation Various tasks are taken place in a substation [10]; these tasks are implemented with the combination of control panel, relays, light meters, cabling and wiring, switches. Earlier in traditional grid engineers had to monitor the substation manually, that was an endeavor and hardworking task. As we switched to smart grid from the conventional grid system, we got a concept of substation automation. It enables the utility to obtain the data from the smart devices, controlling and automation features within the substation. No longer, the engineers must manually check the breakdown and cause of damage in the substation. All the tasks will be now monitored virtually or remotely, by the utility using the control commands. The exchange of information/data between the distribution components is taken place via serial bus.
4
S. Biswal and D. Asija
2.2 Advanced Meter Reading The electromechanical meters can only measure active or reactive power at a time as it only has one dial, but they cannot measure the other factors which are also equally important. There comes the concept of advanced meter reading that automatically collects the data of consumption from the energy meter. The data collected from the energy then goes to the distribution companies where further tasks take place. AMR technologies include cloud-based communication, mobile and network technologies based on telephone networks. With advanced meter reading, there is an improvement in security with tamper detection in the meters; also, it gives accurate meter reading as compared to electromechanically meters. Further, we will discuss in detail under smart meters.
2.3 Advanced Meter Infrastructure As we read about the AMR, advanced meter infrastructure enables to do the operations and control the smart metering system. It enables the utilities to offer new and advanced technology to the consumers to reduce the consumption and peak energy demand [11]. The AMI has four main blocks to focus on: i. ii. iii. iv.
Smart meters installed. Communication network enabled between the consumers and the DISCOMs. Meter data acquisition system (MDAS), collection of data from the meters. Meter data management system (MDMS), to store the information as well as to process it, billing of energy consumption etc.
We can see this in Fig. 2 [11] how this structure or block is basically working; first, the smart meters are installed, and a communication is established between the utility and end users. Now, the data is collected from the meters and then it is managed by the meter data management system; this collectively describes how a two-way communication is established, and relevant data is being processed and controlled by DISCOMs.
Smart Meters
Communication
Fig. 2 Structure of advanced meter infrastructure [11]
Meter Data Acquisition System
Meter Data Management System
Smart Grid and Net Metering Impact in Current …
5
2.4 Distribution Automation Distribution automation (DA) brings developments to the distribution network which functioned autonomously by responding to the signals and about operations. The components which are involved in substation are controlled by DA also control on feeders (known as feeder automation) and components in the meters. Through DA, the utility can collect the data from the smart devices; it automates, optimizes and analyzes them, so that overall efficiency, speed and accuracy of power grid increase. There are lot of applications enabled by DA; it monitors and analyzes the software that controls the whole smart grid system as well as helps in identification of any theft and preventive equipment maintenance.
2.5 Demand Response The demand of electricity in India is high, and to manage it, we need to make a strategic plan to make use of the energy generated so that every house is electrified. This is called as load shifting; this enables the utility to shift energy consumption during peak hours of the day to the other hour 6 s when demand is quite less as compared [12]. The following figure shows the strategies of demand response, i.e., direct load control and load pricing methods. Direct load control basically includes remote interruption of consumer’s energy usage for different activities during off and on peak hours of the day, whereas load pricing is when the utility uses a certain pricing method for different hours of the day (Fig. 3).
Fig. 3 Strategies of demand response [12]
6
S. Biswal and D. Asija
2.6 Energy Management System An energy management system comprises a software-based system which operates the electric grids, so that it can control, monitor and optimize how generation and transmission of energy are taken place in grid. For this purpose, SCADA—real-time data-based computer software—is used for gathering and analyzing. We need this software in order to manage the data as it is lot to cover. This software is also termed as SCADA/EMS. It collects data which located remotely, so that it can transmit it to the utility for controlling and monitoring purpose.
3 Smart Meters As we discussed for the future of a developing country, “smart grid” plays an essential role in that. The most important component in smart grid is “smart meters”; these bought a drastic change to the metering system. These meters use the digital technologies which allows the utility and consumers to have a two-way communication, which was not present in our conventional meters. This meter allows the utility to check the consumption of energy by the consumer over given period of a time. Not only it is beneficial to the utility but also to the end users; the consumers can see how much energy they have used or are using on real-time basis; it helps them to keep check on their consumption and can control it as well by opting reduction in electricity during peak hours of the daytime. As the demand for electricity is reduced, consumer saves a lot of money and put less pressure on power grid. In India as of Feb 2020 Energy Efficiency Services Ltd (Fig. 4). Delhi, Uttar Pradesh, Haryana and Bihar are the states where smart meters have been installed successfully [14].
3.1 Advantages of Smart Meters Advancement in current billing system. Taking an example of electricity connection five years ago, in our homes, we had only 1 AC, and the billing of electricity consumption was based on two-tariff method [15]. So, what DISCOMs do is it will give you 2.5 KW sanctioned load in your homes, so you have to pay a demand charge for it. Now, after five years, suppose some of the consumers have two more ACs in their houses that means three ACs so the load which was sanctioned five years ago will now go up from 2.6 KW to 7 KW [14]. As per the two-tariff method, the fixed charges keep on increasing which further increases the revenue. Billing Efficiency. The conventional meters are electromechanical meters which have disks and dials, the billing was done manually, earlier mainly a person from the distribution companies take the readings from the meter and then the billing of
Smart Grid and Net Metering Impact in Current …
7
Fig. 4 Smart meter-HPL smart meter manufactures [13]
the electricity consumption is generated. Sometimes, many people used to face a problem with the older meters that many meters were dysfunctional also, so what DISCOMs used to do is they bill them using the flat rates. Smart Meter Transparency. Earlier, we had to wait for the bill to get generated so that we can pay [16], but with the help of smart meters there is transparency in the meters, as it shows data time to time the consumers have to no longer wait for 1 month or so for the bill, they can pay their bills on time. The energy consumption by the consumers is shown on real-time basis, and the data is processed within milli sec so it is available virtually on their smart phones; this is done using an app recommended by the distribution companies. Renewable benefits from Smart Meters. Using renewables, power generation happens mostly in daytime. With the “time of the day” policy, different rates for the units consumed are applied at different times of day [17]. Rate of electricity consumed per unit hour is kept low during off-peak hours when demand is not high [18]. It helps the consumers to save lot of money, and it also helps the distribution companies as there is reduced pressure on its grid during peak hours. This can only be done using real-time flexibility; consumers can consume electricity at cheaper rates. This encourages people by demand response as we discussed earlier. The main agenda is that to shift the load to another time [17]; the cost of production of energy is higher as compared to returns generated [15].
8
S. Biswal and D. Asija
3.2 Methodology Adopted by EESL Almost all the distribution companies are working on a combined dashboard; i.e., all the information about the units consumed and trends are stored in that dashboard. The collected data in the distribution company system then follows up a legacy billing software, so that a bill is generated for the consumers [19]. The distribution companies are working on a virtual framework cloud-based. Experts recommended that GPRS [20] network is better than other technologies. They said so because of the costing for the setup, taking an example of radio frequency-RF (narrow band Internet of things) is heavy on cost, it will cost about Rs. 6000 per meter, and when we compare to the GPRS network, it is cheap, and it costs about Rs. 2500 per meter. Each meter is installed with a sim card unit which then transmits the data to the telecom network to the distribution companies’ main server. And for the RF network, an institutional capacity must build so that it can be maintained, that increases the total infrastructure expenditure. About 98.5% GPRS network is available in India, so it will be effort-less transmission of data. EESL also came up with the idea of “combination of network technologies,” by combining narrow band Internet of things (IoT) and GPRS network; RF can be used in multi-story building as the data from all meters can be collected to a single digital control unit, and from DCU, all the information/data collected will go to GPRS network, which will further go to telecom network to the DISCOMs.
3.3 Components Inside a Smart Meter There are three major components [21] of a smart meter. Power Supply. It is provided with a switch-mode power supply with a battery backup so that the metering components/devices stay on even when there is a power cut or main line disabled. This done with the conversion of alternating current voltage to direct current voltage. When there is a power cut, the switch will be on the batter back up, and the conversion takes place. This battery backup is not operated during normal operation of meters. Microcontrollers. Designing a microcontroller is the important step of designing the smart meters; there are several ways in which the smart meter architecture can be designed. • Two chips: It provides with system upgrades. • Single chip: Combination of both hardware and software, which makes it less flexible as compared to two-chip method. • Transmission over a network: Provides with a best communication method which varies with the location.
Smart Grid and Net Metering Impact in Current …
9
Table 1 Types of network and associated equipment [21] Network
Equipment
RF network
Bridges, wireless gateways, relays, access points, routers
Cellular/Telecom
Cell relays, access point
Broadband over Powerline
Gateways, bypass devices
WiMAX
Routers, access points
Communication Interfaces. As we saw, there is no single method for the communication between the smart meters, distribution companies and consumers. Communication protocols vary from place to place; we cannot adopt one single method for all the locations, and it is depended on various factors. Some countries adopted the method of wireless communication over RF signals; this can establish a communication between the utility and the consumers; also, these signals are used in home automation such as Zigbee [21]. The following table shows the network and the associated equipment (Table 1).
4 Grid Stabilization Using PMU—“Phasor Measurement Unit” Stability of smart grids is threatened, and for that matter, we must look for the solutions that can stabilize the smart grids [22]. The following factors lead to destabilized power grid the demand and supply which works on real-time basis. • • • •
Non-inertial and time-variant sources The load connected with the power grids needs a source The complex and variable infrastructure As it is real-time-based network, the systems should have power on.
PMU is basically used for the estimation of the magnitude and phase angle of an electrical phasor quantity in the power grid using a time synchronization. This time synchronization [23] is equipped with a GPS network, which allows the synchronized real-time measurements of the parts of power grid. SCADA is also a real-time-based system that gathers the data and analyze it; like PMU, it also monitors and controls the power plant system and electronic devices in industries. We need such systems to analyze and control, because the data which is gathered from the devices is in bulk which is not feasible for humans to keep a track and analyze at the same time [24]. The system is enclosed so the data transfer is between a SCADA system (computer), programmable logic controller (PLCs), remote terminal units (RTU) and operator terminals [25] (Table 2).
10
S. Biswal and D. Asija
Table 2 Comparing basic differences between RTU-based SCADA and PMU-based SCADA [25] Measurement type
RTU-based SCADA
PMU-based SCADA
Analog based
Digital based
Control and monito ring
Locally
Wide area
Measurement of phasor angle
Not there
Can be measured
Resolution
2–3 measurements per second
Can take up to 50–60 measurements per second
Type of observability
Observability in steady state
Observability in dynamic state
5 Net Metering Earlier, we talked about the intermittent renewable sources like solar energy; solar power is present in abundant. Solar energy is generated when sunlight hits the modules on a solar PV installation module; it thus converts the solar energy into electricity that can be used power various industries and residents [26]. The power from the distribution company and power generated from solar energy that comes to our home do not affect the power performance. The DC power that is produced using solar panels is wired into an inverter, so that it can convert the DC to AC that is required by most of the electrical devices to power on them. This converted AC power from the inverter goes to the service panel in our homes which feeds the electricity to the homes. Smart meters installed at our places keep the track of the total electricity consumed and electricity generated through the installation; it keeps a track online to see how much the consumption is there and whether the electricity demand is met or not also keeps a track of theft [26]. It is a method in which residents and commercial consumers can generate their own electricity from PV solar power. They consume the generated energy and the energy which is not consumed is then returned to the power grid so that they can use that unused energy to any other industries. Customers are billed for the net use; on an average, the customers are billed about 20–30% of solar energy generated [15]. Revenue payback related to the cost of infrastructure is generated which eventually turns into a profit. During any time of the day, if the houses use max electricity than a solar energy produced, then the extra energy is harnessed from the grid as it normally would be. When we opt for this method, we see a difference in our electricity bills, as this installation generates more power and pays for itself overtime, as it harnesses the solar power energy which is abundant to us. Where there is a roof, the utility reduces the need of expensive polluting power plants, giving the world cleaner and pure air to breathe [4]. Sometimes, the solar panels generate more power that is excess as compared to the consumption; this surplus energy then goes to the grid. Other months it might happen that the energy generated by the solar panels is not enough for the whole industry/houses, and then they harness the energy from the grid. So, these two scenarios end up us with either a charge or a credit that for the month. The normal
Smart Grid and Net Metering Impact in Current …
11
rate of the credits the consumers accumulate in a given month can offset the charges they received in other months. Every month the consumer receives the bill statement that reconciles the credits and charge received over the past year.
5.1 Types of Net Metering The following figure shown is with the reference [27] study of the types of net metering and brief about how it works for a single owner and multiple owners (industries) (Fig. 5). Community Net Metering. In community metering [27], multiple consumers form a community. Instead of treating them as multiple consumers, the whole community is treated as a single consumer. They basically set up a common generating solar energy site in the aegis of net meter paradigm rather than setting up solar panels for every single consumer. This type of net metering method is quite like microgrid system, but this is connected to main power grid with some policy rules and regulations. In this type of net metering, consumers may be located at adjacent or non-adjacent sites, and everyone has a separate smarts meters installed which have tariff rates different for everyone; this can be connected to a community shared renewable energy to offset their electricity demand. The main advantage of community net metering is that they save the expenditure in a large scale. As the multiple
Fig. 5 Types of net metering [27]
12
S. Biswal and D. Asija
consumers count in as a single consumer, inconvenience caused by filling up various forms, the operation and maintenance of the power plant and solar panels would be reduced. Aggregate Net Metering. This [27] is another method that allows a consumer having multiple meters in an adjacent location to offset the energy using a single solar energy panel. Such consumers are those who won schools, universities, factories, etc. The consumers have to combine their loads from their meters so that system size can be decided. Only the kWh component of the bill will be credited to the consumers. There is no payment of bills for individual installations; they are benefited with the large-scale economics. Such method of net metering can create new job opportunities and new market for intermittent sources. A single consumer owning multiple meters and planning of allocation of units to different meters would help them to manage the peak load, which results in reduction in peak hour costs that earns them a lot of profit. But with this method, there are some disadvantages too, to determine which sector is eligible for this net metering method and adjacent location for the same. Credits allocation and applied tariff rates need to be designed again. Virtual Net Metering. In virtual net metering system [27], the consumers do not have to install their own solar panels; they gain some benefits from the solar energy. Virtual net metering is applied when commute consumers share on site solar energy or off-grid solar energy [27]. The solar energy system can be shared when it is located at a common place of a large custom solar farm who have about 100 of subscribers. Consumers receive their electricity bill based on the amount of energy generated by their shares of community PV solar installation. In virtual metering, the solar panels are not connected to the smart meters of the subscribers. The solar installation does not provide direct power to the consumers; instead, all the electricity generated goes straight into the grid in return for the credits. The subscribers cannot have the direct power their houses with the solar energy and thus generates high electricity bills.
6 Conclusion In India, the total installed renewable energy capacity is 114.4 GW, i.e., almost 33% of total energy capacity of the country. As the economic developments and industrialization happen in country rapidly, the energy consumption increases. It is predicted that this energy demand will soon outpace the production or generation of energy. All the other sources from where energy is harnessed like coal, natural gases etc. are quite difficult to obtain, and such sources are now shrinking. In such situation, renewable energy sources will soon surpass the natural gas, coal to become the second largest source of energy. One of the steps that is included is net metering initiatives; this initiative not only gives us green energy but also helps people to generate their own energy, saving money and a step toward smart city. Everything around is converting to smart appliances; our homes are now operated with our phones
Smart Grid and Net Metering Impact in Current …
13
known as home automation, and the cars we drive are now using speech recognition just following our commands and performing [28]. The use of smart phones made it possible to overcome the limitations that are found in the smart meters.
7 Future Scope Earlier, we discussed that Indian Power System is facing problems related to revenue; the infrastructure of our grid was degrading day by day. In future, in order to make our grid system more efficient and smarter, we can adopt the RF technology along with artificial intelligence [29]. The world is changing, and we need to adopt changes and acquire skills for the same. The revenue that is generated and will be generated in future can be used to make our technology more advanced so that the two-way communication in smart grid can be faster, and with that technology, we get security, i.e., hardware security; the data transmission in RF technology is faster and also does not stop its operation even in outages, whereas in GPRS network, there are chances [28]. With the technology so advancing, we need artificial intelligence in our power system network; humans can monitor on real-time basis, but with the help of AI, it makes it easy for the people as the word says Artificial Intelligence [29, 30]; it will decide what needs to be done by closely monitoring (real-time data); for such system and technology, we require RF technology; as for India, it will take ten years from now to adopt this method and amending the changes.
References 1. Jha IS, Sen S, Kumar R (2014) Smart grid development in India—a case study. 2014 Eighteenth national power systems conference (NPSC), Guwahati, India, pp 1–6. https://doi.org/10.1109/ NPSC.2014.7103866 2. Zhou J, He L, Li C, Cao Y, Liu X, Geng Y (2013) What’s the difference between traditional power grid and smart grid?—from dispatching perspective. 2013 IEEE PES Asia-pacific power and energy engineering conference (APPEEC), Kowloon, pp 1−6. https://doi.org/10.1109/APP EEC.2013.6837107 3. Mavridou A, Papa M (2011) A situational awareness architecture for the smart grid. in: Global Security, Safety and Sustainability & e- Democracy, vol 99, Georgiadis CK, Jahankhani H, Pimenidis E, Bashroush R, Al-Nemrat A (eds) e-Democracy 2011, ICGS3 2011. Lecture notes of the institute for computer sciences, social informatics and telecommunications engineering. Springer, Berlin, Heidelberg 4. Rezgui J, Cherkaoui S, Said D (2012) A two-way communication scheme for vehicles charging control in the smart grid. 2012 8th international wireless communications and mobile computing conference (IWCMC), Limassol, pp 883−888. https://doi.org/10.1109/IWCMC. 2012.6314321 5. Shahram Javadi, Shahriar Javadi (2010) Steps to smart grid realization. In CEA’10: proceedings of the 4th WSEAS international conference on Computer engineering and applications, pp 223–228 6. https://powermin.nic.in/en/content/generation-capacity
14
S. Biswal and D. Asija
7. Anuradha M Annaswamy, Massoud Amin (2013) IEEE vision for smart grid controls: 2030 and beyond. In IEEE vision for smart grid controls: 2030 and beyond, pp 1–168. https://doi. org/10.1109/IEEESTD.2013.6577608 8. https://www.electricalindia.in/smart-grid-empowering-smart-city/ 9. Luis I Minchala- Anila, Jairo Armjjo, Daniel Pesantez, Youmin Zhang (2016) Applied energy symposium and forum, REM2016: renewable energy Integration with mini/microgrid. Design and implementation of a smart meter with demand response capabilities, Maldives [online]. https://doi.org/10.1016/j.egypro.2016.11.272 10. https://www.electricalindia.in/substation-automation/ 11. Selvam C, Srinivas K, Ayyappan GS, Venkatachala Sarma M (2012) Advanced metering infrastructure for smart grid applications. 2012 international conference on recent trends in information technology, Chennai, Tamil Nadu, pp 145–150. doi: https://doi.org/10.1109/ICRTIT. 2012.6206777 12. Pierluigi Siano (2014) Demand response and smart grids—a survey. Renew Sustain Energy Rev Elsevier 30:461–478. https://doi.org/10.1016/j.rser.2013.10.022 13. https://www.hplindia.com/product/product-detail.php?id=35 14. https://scroll.in/article/955346/interview-india-is-losing-rs-100000-crore-in-unbilled-electr icity-the-solution-is-smart-meters 15. Tony T, Sivraj P, Sasi KK (2016) Net energy meter with appliance control and bi-directional communication capability. 2016 international conference on advances in computing, communications and informatics (ICACCI), Jaipur, pp 2650–2653. https://doi.org/10.1109/ICACCI. 2016.7732458 16. Tony Flick, Justin Morehouse (2011) Threats and impacts: utility companies and beyond. In Securing the smart grid, Tony Flick, Justin Morehouse, Syngress, pp 35–48, ISBN 9781597495707, [Online]. https://doi.org/10.1016/B978-1-59749-570-7.00003-0 17. Logenthiran Thillainathan, Srinivasan D, Vanessa K (2014) Demand side management of smart grid: Load shifting and incentives. J Renew Sustain Energy, 6. https://doi.org/10.1063/1.488 5106 18. Pierluigi Siano (2019) Beyond technical smartness: rethinking the development and implementation of sociotechnical smart grids in India. Energy Res Soc Sci Elsevier 49:158–168. https:// doi.org/10.1016/j.erss.2018.10.026 19. Ramakrishna Kappagantu, Arul Daniel S (2018) Challenges and issues of smart grid implementation: a case of Indian scenario. J Elect Syst Inf Technol. Elsevier 5(3):453–467. https:// doi.org/10.1016/j.jesit.2018.01.002 20. Sultan Z, Jiang Y, Malik A, Ahmed SF (2019) GSM based smart wireless controlled digital energy meter. 2019 IEEE 6th international conference on engineering technologies and applied sciences (ICETAS), Kuala Lumpur, Malaysia, pp 1–6. https://doi.org/10.1109/ICETAS48360. 2019.9117479 21. Mike Smith (2021) Monitoring and protecting smart meter circuitry and communications. Bourns 22. Gupta DK, Pandey RK (2014) Grid stabilization with PMU signals—a survey. 2014 eighteenth national power systems conference (NPSC), Guwahati, India, pp 1–6. https://doi.org/10.1109/ NPSC.2014.7103812 23. https://en.wikipedia.org/wiki/Phasor_measurement_unit 24. Annor-Asante M, Pranggono B (2018) Development of smart grid testbed with low-cost hardware and software, for cybersecurity research and education. Wireless Pers Commun 101:1357–1377. [online]. https://doi.org/10.1007/s11277-018-5766-6 25. Bentarzi H, Tsebia M, Abdelmoumene A (2018) PMU based SCADA enhancement in smart power grid. 2018 IEEE 12th International conference on compatibility, power electronics and power engineering (CPE-POWERENG 2018), Doha, pp. 1–6. https://doi.org/10.1109/CPE. 2018.8372580 26. Kavita Narayan Mule, Prof Deshmukh BT (2018) Design of smart energy meter for smart grid with net metering & theft detection. Intern J Res Eng Appl Manag (IJREAM) 03(12). Aurangabad(MS), India. https://doi.org/10.18231/2454- 9150.2018.0068
Smart Grid and Net Metering Impact in Current …
15
27. Thakur J, Chakraborty B (2015) Smart net metering models for smart grid in India. 2015 International conference on renewable energy research and applications (ICRERA), Palermo, pp 333–338. https://doi.org/10.1109/ICRERA.2015.7418720 28. Dileep G (2020) A survey on smart grid technologies and applications. Renew Energy, Elsevier 146:2589–2625. https://doi.org/10.1016/j.renene.2019.08.092 29. Muhammad Qamar Raza, Abbas Khosravi (2015) A review on artificial intelligence based load demand forecasting techniques for smart grid and buildings. Renew Sustain Energy Rev 50:1352–1372, Elsevier. https://doi.org/10.1016/j.rser.2015.04.065 30. Khoussi S, Mattas A (2016) A brief introduction to smart grid safety and security. In Handbook of system safety and security, pp 225–252 [online]. https://doi.org/10.1016/B978-0-12-8037737.00011-5
Machine Learning-Based Approach for Myocardial Infarction Pooja Maindarkar and S. Sofana Reka
Abstract In this work, machine learning-based classifiers for the detection of a heart disease called as myocardial infarction (MI), commonly known as heart attack, are done. It is dominantly observed that myocardial infection causes rapid increase in mortality rate day by day, and thereby, it is necessary to develop a model which is efficient and reliable. In this work, four popular noninvasive machine learning classifiers, namely SVM, K-NN, logistic regression and random forest, were involved for testing and evaluated the results of these classifiers on ECG signals obtained from PhysioNet (PTB diagnostic ECG dataset). The classifiers were based on using performance evaluation of the model such as accuracy, sensitivity, specificity, area under curve and processing time. On comparing, the four classifiers used support vector machine (SVM), random forest, K-NN and logistic regression. Logistic regression outperforms all the other classifiers with 88% accuracy, 93% sensitivity, 82% specificity. The result on comparison demonstrates how different machine learning classifiers respond to the ECG dataset in terms of accuracy, specificity and sensitivity. Keywords K-Nearest neighbor (K-NN) · Myocardial infarction (MI) · Support vector machine (SVM) · Electrocardiograph (ECG)
1 Introduction Globally, myocardial infraction (MI) is considered to be one of the most dangerous diseases. According to the statistics available till 2018, it has been observed that 17.9 millions of deaths that occur worldwide are due to the heart disease [4]. As MI is one of the important reasons for rapid increase in the mortality rate, prevention and detection at early stage of the disease are very much required in the health perspective at the larger scale. In clinical language, heart attack is called myocardial infarction. In MI, the nutrient and oxygen supply to the heart muscles decrease. Due to the deposition of plaques in the heart veins, the veins get blocked and the further supply of P. Maindarkar · S. S. Reka (B) School of Electronics Engineering, Vellore Institute of Technology, Chennai, Tamilnadu, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_2
17
18
P. Maindarkar and S. S. Reka
oxygen and nutrients through vein to heart muscle cells stops, resulting in the death of the patient [1]. Heart disease can be diagnosed through various medical tests like blood test and echocardiogram (ECG). Till now, ECG has proven to be effective in determining the MI. It is noninvasive and inexpensive. The traditional way of curing the disease is not reliable; hence for the accurate detection and prevention, we require a computer-aided system which is more efficient [2]. In developed and underdeveloped countries, the detection of MI is a challenging task. Hence, various researchers have proposed different computer-based smart systems for the accurate detection of the heart failure [7]. One of the approaches mentioned in this work [27] exhibits a hybrid system with machine learning classifiers J48, Naive Bayes and random forest. Among these three identifiers, random forest classifier has shown the accuracy of 100%. It has been observed that the dataset feed to these classifiers contains many irrelevant features. These features affect the performance and accuracy of the classifier; hence, it is necessary to use the feature selection methods to remove the less contributing features to the system. In this work, authors [28] have used relief feature selection algorithm, minimal redundancy maximal relevance, selection operator and least absolute shrinkage algorithms for the feature selection purpose. After removing the irreverent features from dataset, it is trained through seven different classifiers such as NB, ANN, SVM, DT, K-NN and logistic regression in this work. And it is observed that SVM (RBF) and logistic regression show the best accuracy of 88% and 89%. Two SVM classifiers are stacked [22]; the first SVM performs the work of eliminating the irrelevant features making their coefficients to zero, and the second SVM predicts the heart failure detection. This model helps to improve the efficiency by 3.3% compared to other conventional SVM models. Similarly in this work, the authors [26] have proposed an automatic heart failure detection system using ECG signal using SVM and its duality for the maximum efficiency is done. Here, the output of SVM classifier is compared with its duality for greater performance. The accuracy of this system is 94.97% [26]. Some of the methods that have been implemented recently in machine learning face the problem of over-fitting, the disease detection accuracy is improved on testing data, but the accuracy of training data is compromised. To overcome this problem, two machine learning algorithms, random search algorithm (RSA) for different feature selections and random forest classifier for cardiac failure prediction [23], are done by the authors in this work. Both these algorithms are dependent on each other and execute one after another making the system hybrid and more efficient. The accuracy is found to be 93.33%. In clinical data mining, it is found that the unstructured data is heterogeneous in nature. For accurate result of heart disease, it is important to study the various attributes of the data which is uncertain in nature. Hence, to remove the uncertainties various data mining techniques are implemented but they are found to be in accurate. In the proposed work [16], a fuzzy K-NN model is proposed by the authors where the standard deviation and mean along with an exponential function are calculated for each person and this value is then multiplied with the attribute to remove the uncertainty and make the model more efficient. In the literature, numerous works have been done related to concerned perspective.
Machine Learning-Based Approach for Myocardial Infarction
19
2 Materials and Methods In this work, the required involved by testing the output of the technique on ECG signals from Physionet (PTB diagnostic ECG dataset) [2]. It is an open-access dataset. Signals are taken from the groups of subjects with respect to the normal group and the MI-affected group. The normal group contains 80 healthy subjects, and the patient group contains 80 MI-affected subjects. Each signal is in standard format with 10 s as sample size. The dataset is acquired simultaneously having 15 leads including a 3-lead Frank VCG and traditional 12 leads. The leads such as i, ii, iii, aVR, aVL, aVF and V1-V6 are 12-lead ECG signal, and the remaining three leads (Vx, Vy and Vz) are vector cardiogram. Also to gain more accurate results and enhance input features, statistical parameters such as mean median and standard deviation were also included during the training process. Table 1 exhibits the details of 12 leads of ECG dataset. In the proposed work, MI is one of the major concerns for the maximum number of death worldwide. Many related works in the literature have been proposed by various researchers to detect and analyze this disease at early stage through various methods, such as analyzing MR image sequences and using wearable medical sensors and machine learning. In this paper, we propose to detect MI through four different machine learning algorithms using ECG signal parameters as input features. The proposed model of each algorithm is capable to classify the person as healthy or affected with more than 85% accuracy in all the four models with comparison done in the study. On observing the obtained dataset, the work was initiated with supervised learning algorithms as the output target for each subject/person was clearly defined. Hence, by using well-known supervised algorithms we can compare and analyze which algorithm best suits the given type of ECG dataset. Classification of these subjects as healthy or affected is successfully evaluated through four classifiers, respectively. These classifiers are, namely, logistic regression, random forest classifier, support vector machine and K-nearest neighbor. Initially, the complete dataset is randomly divided by the processing unit as 80% training data and 20% test data. As the training dataset is obtained, it is used for training the model, where the weights are adjusted internally in the model as per the patterns fed as input to this Table 1 ECG dataset attributes Category
Leads
Activity
Inferior
Leads II, III, lead aVF
It looks at the electrical activity process from the vantage point—inferior surface
Lateral
Leads I, Avl Lead V5 , lead V6
Electrical activity from the vantage point—lateral wall of the ventricle left
Septal
V1 V2
Electrical activity from the point of vantage of the inter-ventricular septum
Anterior
V3 V4
Electrical activity from the point of vantage of the anterior wall of the both right and left ventricles
20
P. Maindarkar and S. S. Reka
model. As the training is done, then trained model is tested using the test dataset, the obtained output is verified with the original output, and the accuracy is calculated. In machine learning, K-NN is found to be the most basic and essential classification algorithms. It is supervised learning algorithm and has various applications in pattern recognition, intrusion detection and data mining. K-NN is nonparametric, and in real-time scenarios, it is widely disposable where it exhibits about the distribution of data as it does not make any underlying assumptions as opposed to other algorithms such as GMM, which assume a Gaussian distribution of the given data. Hence, this algorithm is selected for the detection of MI using ECG signal. The support vector machine (SVM) is another ML algorithm which is capable of performing classification tasks using classifiers in a multi-dimensional space which separates the data into two different classes and can support both classification tasks and regression and can handle multiple categorical and continuous variables. SVM algorithms make use of mathematical functions that are mainly defined as kernel. The major function of a kernel is to accept data as an input and change it into the desired form. Various SVM algorithms have different types of kernel functions, such as nonlinear, polynomial, linear, radial basis function (RBF) and sigmoid. Kernel functions return back the innermost product between two points in a suitable feature. For the training purposes involved, after evaluating for all type of kernels for modeling, the linear kernel was used, which defines that the dataset has linearity property among the parameters to predict the output. Logistic regression is also known as the statistical model that basically makes use of logistic function to model the variable which are binary dependent. In analyzing regression, logistic regression estimates the parameters of a logistic model. In this model which uses sigmoid function which converts a linear line to a curve to cover the binary points. This model is very useful to find probabilities of a certain event, and it can also be used to find whether a given data will fit into either of the binary values (i.e., 0 or 1). y = b0 + b1 x p= ln
p 1− p
1 1 + e−y = b0 + b1 x
(1) (2) (3)
Random forest algorithm is almost similar to decision tree algorithm. The only difference in random forest is it has multiple decision trees inside single algorithm. The final decision of classification is made by the maximum number of trees. This algorithm is selected in this work majorly through the literature survey; it was found that this algorithm reduces the risk of over-fitting with required training time and claims to offer high accuracy for classification. It also claims to provide high accuracy in case of missing data. And the necessity of such algorithm is there in the applications such as detecting heart disease of a person. Some important terms which affect the model are entropy, gain, leaf node, decision node, root node.
Machine Learning-Based Approach for Myocardial Infarction Table 2 Confusion matrix
21
Category
Predicted MI patients (1)
Predicted Healthy patients (0)
Actual MI patients (1)
TP
FN
Actual healthy patients (0)
FP
TN
The performance of proposed work is evaluated through various parameters like confusion matrix, accuracy, classification error, sensitivity, specificity and precision which are estimated in accordance with Table 2. Confusion Matrix: It gives us the insight into how the classifier got confused while making the prediction. The numbers of correct and incorrect predictions are summarized in two classes. It is a 2 × 2 matrix. Table 2 shows the confusion matrix. Here, 1 represents MI-affected case and 0 represents the not affected, normal case. Accuracy: It shows the overall performance of the model. And it can be calculated Accuracy =
TP +TN T P + T N + FP + FN
× 100%
Classification Error: It is the overall incorrect classification of the classification: Err or =
FP + FN T P + T N + FP + FN
× 100%
Sensitivity: It mainly measures the part of actual positives that are correctly predicted as such (e.g., the percentage of affected person who is actually predicted as having the condition). Sensitivit y =
TP T P + FN
× 100%
Specificity: It mainly measures the part of actual negatives that are correctly predicted as such (e.g., the percentage of healthy people who are correctly predicted as not having the condition). Speci f icit y =
TN T N + FP
× 100%
Precision: It is defined as measure of positive predictive values that are actually positive
22
P. Maindarkar and S. S. Reka
Pr ecision =
TP T P + FP
× 100%
3 Results and Discussion In this section of work, the evaluation of the outcomes from the four classification systems, namely random forest, logistic regression, K-NN and SVM, has been done. The work done here shows how these machine learning classifiers have shown different responses on the PTB diagnostic ECG dataset. It is observed that the logistic regression classifier outperforms the other 3 classifiers with 88% of precision. In our experiments with K-NN classifier, we performed the modeling for various values of k = 1, 2,3,4,…19. The performance was observed comparatively good at k = 8 with the accuracy of 82% as compared to the other values of k, like for the value k = 2 the accuracy is only 45%. Figure 1 shows the performance of K-NN on different k values, and Table 3 displays the result matrix. The performance of random forest and support vector machine (kernel = linear) is observed to be same as 82%. In random forest classifiers, 10,20,200,500 and 1000 iterations were applied to get the best precision, and among all the iterations, the classifier has good accuracy of 84% at minimum iteration of 10. In random 1.
1.
Fig. 1 Accuracy and classification matrix
Table 3 K-NN result matrix
Category
Precision
Recall
F1 score
0
0.76
0.95
0.84
1
0.93
0.70
Accuracy
0.80 0.82
Machine Learning-Based Approach for Myocardial Infarction Table 4 Random forest result matrix
23
Category
Precision
Recall
F1 score
0
0.94
0.79
0.86
1
0.75
0.92
0.83
Accuracy
0.84
1.
1.
Fig. 2 Accuracy and classification matrix for RFA
forest, it can be seen that as the number of iterations is increasing from 10 to 1000 the performance of classifier goes down. Table 4 displays the result matrix for the random forest algorithm with 10 number of trees. And Fig. 2. shows the confusion matrix for the same. Also in Fig. 2, comparison of accuracies is shown with respect to number of trees formed. While performing support vector machine algorithm, all the kernels were tested and evaluated and the best results were obtained through linear kernel. Table 5 displays the result matrix for linear kernel SVM, and Fig. 6 shows the confusion matrix for the same. At last, the model was trained using logistic regression and the accuracy was found to be 88%, which is highest among all the other classifiers which were test along with good precision, recall and F1 score. Table 6 represents the result matrix for logistic regression model, and Fig. 3 represents the classification matrix for the same. After observing all the resultant parameters, it is clearly understood that among these classifiers, logistic regression shows the best performance and can be applied to the given type of dataset of ECG signal for various people. From Table 7, it can be concluded that logistic regression outperforms the classification and with Table 5 Linear SVM result matrix
Category
Precision
Recall
F1 score
0
0.89
0.84
0.86
1
0.79
0.85
0.81
Accuracy
0.84
24
P. Maindarkar and S. S. Reka
Table 6 Logistic regression result matrix
Category
Precision
Recall
F1 score
0
0.82
0.93
0.87
1
0.93
0.82
0.87
Accuracy
0.88
1.
Fig. 3 Classification matrix for linear SVM and logistic regression
Table 7 Performance of classifiers Model
Accuracy (%)
Sensitivity (%)
Specificity (%)
AUC (%)
Processing time (sec)
Logistic regression
1. 88
1. 93
1. 82
1. 87.84
1. 0.09
Random forest
2. 84
2. 79
2. 92
2. 85.62
2. 0.20
Linear SVM
3. 84
3. 84
3. 85
3. 79.41
3. 0.10
K-NN
4. 82
4. 95
4. 70
4. 82.5
4. 0.50
less processing time compared to other classifiers. Figure 4 displays the overall comparison plot of all the algorithms for three major parameters which are accuracy, sensitivity and specificity.
Machine Learning-Based Approach for Myocardial Infarction
25
Fig. 4 Performance of classifiers
4 Conclusion and Future Scope In this paper, we have compared and evaluated different classifiers using ECG signal information from PhysioNet (PTB diagnostic ECG dataset). All these models were compared on various resultant parameters such as sensitivity, specificity, precision and accuracy. It was also verified that processing time for logistic regression is comparatively very less compared to other classifiers. Hence, it can be applied over clinical dataset and patients can be testified successfully. In the near future, ensemblebased logistic regression can be performed to improvise the accuracy of the algorithm even further. Also, the model can be developed using logistic regression by training with images of the affected and healthy subjects, as this model depends of probabilities; the class having maximum probability is selected as the final output class, and images can provide more input features and improvise the decision making of an algorithm.
References 1. Chen J, Valehi A, Razi A (2019) Smart heart monitoring: early prediction of heart problems through predictive analysis of ECG signals. IEEE Access 7:120831–120839 2. Sopic D, Aminifar A, Aminifar A, Atienza D (2018) Real-time event-driven classification technique for early detection and prevention of myocardial infarction on wearable systems. IEEE Trans Biomed Circuits Syst 12(5):982–992 3. Liu W et al (2018) Real-time multilead convolutional neural network for myocardial infarction detection. IEEE J Biomed Health Inform 22(5):1434–1444
26
P. Maindarkar and S. S. Reka
4. Behloul F, Lelieveldt BPF, Boudraa A, Janier MF, Revel D, Reiber JHC (2001) Neurofuzzy systems for computer-aided myocardial viability assessment. IEEE Trans Med Imaging 20(12):1302–1313 5. Feng Kai, Pi Xitian, Liu Hongying Sun Kai (2019) Myocardial infarction classification based on convolutional neural network and recurrent neural network. Appl Sci 9:1879. https://doi. org/10.3390/app9091879 6. PU Yi-Fei (2016) Measurement units and physical dimensions of Fractance-part II: fractionalorder measurement units and physical dimensions of Fractance and rules for Fractors in series and parallel. IEEE Access. 4:1−1. https://doi.org/10.1109/ACCESS.2016.2585819 7. Gjoreski Martin, Anton Gradišek, Borut Budna, Matjaž Gams, Gregor Poglajen (2020) Machine learning and end-to-end deep learning for the detection of chronic heart failure from heart sounds. IEEE Access 8:20313–20324 8. Beulah Christalin Latha C, Carolin Jeeva S (2019) Improving the accuracy of prediction of heart disease risk based on ensemble classification techniques. Inform Med Unlocked 16:100203, ISSN 23529148. 10.1016/j.imu.2019.10203. http://www.sciencedirect.com/science/article/pii/ S235291481830217X 9. Baydoun M, Safatly L, Abou Hassan OK, Ghaziri H, El Hajj A, Isma’eel H (2019) High precision digitization of paper-based ECG records: a step toward machine learning IEEE J Transl Eng Health Med 7(1900808):1–8 10. Kampouraki A, Manis G, Nikou C (2009) Heartbeat time series classification with support vector machines. IEEE Trans Inf Technol Biomed 13(4):512–518 11. Jabbar MA, Samreen S (2016) Heart disease prediction system based on hidden naïve bayes classifier. 2016 International Conference on Circuits, Controls, Communications and Computing (I4C), Bangalore, pp 1–5 12. Zhao L, Liu C, Wei S, Liu C, Li J (2019) Enhancing detection accuracy for clinical heart failure utilizing pulse transit time variability and machine learning. IEEE Access 7:17716–17724 13. Kumar SS, Shaikh T (2017) Empirical evaluation of the performance of feature selection approaches on random forest. 2017 International Conference on Computer and Applications (ICCA), Doha, pp 227–231 14. Mohan S, Thirumalai C, Srivastava G (2019) Effective heart disease prediction using hybrid machine learning techniques. IEEE Access 7:81542–81554 15. Sun L, Lu Y, Yang K, Li S (2012) ECG analysis using multiple instance learning for myocardial infarction detection. IEEE Trans Biomed Eng 59(12):3348–3356 16. Krishnaiah V, Srinivas M, Narsimha G, Chandra NS (2014) Diagnosis of heart disease patients using fuzzy classification technique. International Conference on Computing and Communication Technologies, Hyderabad, pp 1–7 17. Strodthoff Nils, Claas Strodthoff (2019) Detecting and interpreting myocardial infarction using fully convolutional neural networks. Physiol Meas 40(1):015001. Crossref.Web 18. Chen M, Fang L, Zhuang Q, Liu H (2019) Deep learning assessment of myocardial infarction from MR image sequences. IEEE Access 7:5438–5446 19. Zhang J et al (2017) Coupling a fast fourier transformation with a machine learning ensemble model to support recommendations for heart disease patients in a Telehealth environment. IEEE Access 5:10674–10685 20. Venu K, Natesan P, Krishnakumar B, Sasipriya N (2019) Classification of myocardial infarction using convolution neural network. Intern J Rec Technol Eng Reg 8(4):12763–12768. https:// doi.org/10.35940/ijrte.d9230.118419 21. Tripathy RK, Dandapat S (2017) Automated detection of heart ailments from 12-lead ECG using complex wavelet sub-band bi-spectrum features. Healthcare Technol Lett 4(2):57–63 22. Ali Liaqat, Awais Niamat, Noorbakhsh Amiri Golilarz, Akhtar Ali, Xiong Xing-zhong (2019) An expert system based on optimized stacked support vector machines for effective diagnosis of heart disease 23. Javeed A, Zhou S, Yongjian L, Qasim I, Noor A, Nour R (2019) An intelligent learning system based on random search algorithm and optimized random forest model for improved heart disease detection. IEEE Access 7:180235–180243
Machine Learning-Based Approach for Myocardial Infarction
27
24. Pawlovsky AP (2018) An ensemble based on distances for a kNN method for heart disease diagnosis. 2018 International Conference on Electronics, Information, and Communication (ICEIC), Honolulu, HI, pp 1–4 25. Tripathy RK, Bhattacharyya A, Pachori RB (2019) A novel approach for detection of myocardial infarction from ECG signals of multiple electrodes. IEEE Sens J 19(12):4509–4517 26. Geweid GGN, Abdallah MA (2019) A new automatic identification method of heart failure using improved support vector machine based on duality optimization technique. IEEE Access 7:149595–149611 27. Dhar S, Roy K, Dey T, Datta P, Biswas A (2018) A hybrid machine learning approach for prediction of heart diseases. 2018 4th International Conference on Computing Communication and Automation (ICCCA), Greater Noida, India, pp 1–6 28. Haq AU, Li JP, Memon MH, Nazir S, Sun R (2018) A hybrid intelligent system framework for the prediction of heart disease using machine learning algorithms. Mobile Inf Syst 2018(3860146) 29. Yin H, Jha NK (2017) A health decision support system for disease diagnosis based on wearable medical sensors and machine learning ensembles. IEEE Trans Multi-Scale Comput Syst 3(4):228–241 30. Islam MM, Xin Yao, Murase K (2003) A constructive algorithm for training cooperative neural network ensembles. IEEE Trans Neural Net 14(4):820–834
Joyful Jaccard: An Analysis of Jaccard-Based Similarity Measures in Collaborative Recommendations Anisha Jindal, Naveen Sharma, and Vijay Verma
Abstract Recommender systems (RSs) are utilized by various e-commerce giants such as Amazon, YouTube and Netflix for providing a personalized experience to the individual users. For this reason, it has become very important to develop and use efficient techniques to provide recommendations. Neighborhood-based collaborative filtering approaches are traditional techniques for recommendations and are very popular due to their simplicity and efficiency. Neighborhood-based recommender systems use numerous kinds of similarity measures for finding similar users or items. Further, defining novel ways to model the notion of similarity is an active thread of research among RS researchers (by using rating data available in the useritem matrix). This work compares and analyzes the performance of various Jaccardbased similarity measures. Primarily, various Jaccard-based similarity measures are examined with their authoritative definitions. We have conducted various experiments using standardized benchmark datasets (MovieLens-100 K, MovieLens-1 M, and Yahoo music) for assessing the performance of these measures. Empirically obtained results demonstrate that the New Weighted Similarity Measure (NWSM) provides better predictive accuracy among all measures. Keywords Recommender systems · Collaborative filtering · Similarity measures · Jaccard index · E-commerce
1 Introduction Collaborative filtering is the most extensively used and traditional recommendation technique [1, 2]. It suggests the items in accordance with the user’s historical preference [3, 4]. Collaborative filtering can be implemented as memory-based or modelbased [5]. The memory-based approach utilizes the user-item ratings to find similar neighbors who share the same tastes with the user and then generate accurate predictions. Here, similarity weight computation has a crucial effect on the performance A. Jindal (B) · N. Sharma · V. Verma Computer Engineering Department, National Institute of Technology, Kurukshetra, Haryana 136119, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_3
29
30
A. Jindal et al.
Table 1 Jaccard-based similarity measures S. No
Measures Name
Reference
1
JMSD
[11]
2
CJMSD
[12]
3
Relevant Jaccard (RJaccard)
[13]
4
JacRA
[14]
5
JacLMH
[15]
6
Cosine–Jaccard–Mean Measure of Divergence (CJacMD)
[16]
7
New Weighted Similarity Measure (NWSM)
[17]
8
New Heuristic Similarity Model (NHSM)
[18]
of the recommendation system. The memory-based approach is easier to implement and can handle new data, but its efficiency reduces with increasing sparsity in data [6, 7]. In contrast, model-based approach uses item ratings to learn a predictive model, and the trained model is then used to make recommendations [8, 9]. It is capable of handling sparsity problems, but this performance comes at the cost of great resources like time and memory. In this work, we aim our attention at neighborhood-based collaborative recommendation approaches. One intrinsic task of these approaches is to discover similar users or items [10]. In the RS literature, the concept of similarity is defined in numerous different ways for collaborative recommendation scenario. Particularly, this work analyzes only the Jaccard-based similarity measures which are effectively used to define the similarity between users/items. Table 1 lists these Jaccard-based similarity measures.
2 Related Work The study of similarity measures, used for collaborative recommendations, has been done in various research articles. The work in [19, 20] compares traditional similarity measures such as Pearson Correlation Coefficient (PCC) [21], Cosine Similarity (COS) [5], Jaccard [22] and Euclidian distance, whereas a few works focus only on the similarity measures available or implemented in a specific library, for instance, the work in [23] assesses the similarity measures available in CF4J library [24] and work in [25] compares the similarity measures in Apache Mahout framework [26]. There exist a lot of works that propose a new way to define the notion of similarity measures in collaborative recommendations [27–30]. For example, research [30] modifies the simple matching coefficient together with Jaccard index for defining a new similarity measure while the work in [29] integrates triangle similarity with the Jaccard coefficient, known as Triangle Multiplying Jaccard (TMJ).
Joyful Jaccard: An Analysis of Jaccard-Based …
31
A recent and noteworthy work reviews the most commonly used similarity measures in collaborative recommendations [31]. Authors have discussed the limitations of the common existing similarity measures and suggested some recommendations that may be considered in the process of computing the similarity in order to improve the overall quality of a recommender system. Another recent work [32] that is closely related to the proposed work compares the similarity measures that akin to the Jaccard index with respect to their definitive formulae. In contrast to the work done in [32], this work analyzes the similarity measures that are based on Jaccard index, i.e., we have provided a comparison of similarity measures which integrate Jaccard index with other existing similarity measures or factors. Jaccard index is the simplest of all existing similarity measures. Its interpretation is quite easy and obvious, but it may provide inaccurate results sometimes. Its simplicity can be exploited to achieve better quality similarity indexes. Various Jaccard-based similarity indexes present in literature are discussed in the next section.
3 Jaccard-Based Similarity Measures 3.1 Jmsd It is the simplest Jaccard-based metric which combines Jaccard and mean-squared difference (MSD) [33] to form the similarity measure [11]. Ix ∩ I y i∈Ix ∩I y JMSD(x, y) = Ix ∪ I y
2 1 − r x,i − r y,i Ix ∩ I y
(1)
3.2 Cjmsd This measure improves the JMSD measure by adding a term corresponding to the coverage that a user u can provide to another user v, i.e., Coverageu→v = |Iu|I−I| v | . Clearly, this similarity measure is asymmetric as Coverageu - > v = Coveragev - > u and hence the similarity between two users, u and v, CJMSD(u,v) = CJMSD(v,u) [12]. C J M S D(u, v) =
|Iu − Iv | × J ac(u, v) × MSD(u, v) |I |
(2)
32
A. Jindal et al.
3.3 Relevant Jaccard (RJaccard) Jaccard measure only incorporates co-rated items in the similarity, ignoring the total length of rated items. On the contrary, it was found that the similarity also depends upon the un-co-rated items of nearest neighbors as well as of target users. RJaccard considers all such relevant factors. If I u is set of items un-co-rated by user u, then RJaccard is defined as below [13]. sim(u, v) R jaccar d =
1 1+
1 |Iu ∩|v |
+
|I u | 1+|I u |
+
1 1+|I v |
(3)
where if |Iu ∩ |v |= 0 then sim(u, v) R jaccar d = 0.
3.4 JacRA Actual ratings given by the users reflect the degree of their preference for the item. Combining both the objective (number of co-rated items) and subjective aspect ( absolute rating values) of a user’s taste, JacRA is defined as follows [14]. |Iu ∩ Iv | sim(u, v) = · |Iu ∪ Iv |
min(rui ,rvi ) i∈Iu ∩Iv max(rui ,rvi )
|Iu ∩ Iv |
(4)
3.5 JacLMH Most of the similarity measures give equal weightage to each rating value. But statistics show that users tend to give ratings higher than the median and avoid extreme values, i.e., some rating values occur more frequently than others. Hence, it is necessary to consider the similarity of extreme ratings and common/middle ratings separately. To formulate it, the rating range is divided into subintervals, and the Jaccard similarity of each of the subinterval is calculated separately [15]. Jac L (u, v) =
|I L ,u ∩ I L ,v | |I L ,u ∪ I L ,v |
(5)
Jac M (u, v) =
|I M,u ∩ I M,v | |I M,u ∪ I M,v |
(6)
Jac H (u, v) =
|I H,u ∩ I H,v | |I H,u ∪ I H,v |
(7)
Joyful Jaccard: An Analysis of Jaccard-Based …
JacLMH(u, v) =
33
1 (Jac L (u, v) + Jac M (u, v) + Jac H (u, v)) 3
(8)
I L ,u , I M,u and I H,u are the set of items having low, middle and high ratings respectively. They are created according to user-defined boundary parameters—H bd and L bd .
3.6 CJacMD It integrates Cosine, Jaccard and MMD. In order to express an individual’s personal habits, mean measure of divergence (MMD) is a powerful tool. θ u represents a vector based on user u ratings, |Iu| represents a total number of ratings made by u and r represents the number of co-rated items between u and u .
sim(u, u )
MMD
=
1+
( r1
1 2 i=1 {(θu − θu ) −
r
1 |Iu |
−
1 }) |Iu |
(9)
C J acM D COS J accar d MMD sim u, u = sim u, u + sim u, u + sim u, u (10) This similarity index is specially designed for sparse environments.
3.7 New Weighted Similarity Measure (NWSM) Neighbors do have an influence on the user. The items that are not rated by the user but by neighbors have a significant impact on his recommendations. Thus, an influence weight has to be associated with each user for every neighbor. Also, the rating criteria for each user is not the same. Some users prefer higher ratings than others. This user behavior is depicted by user rating preference(URP) [17].
|Ib |−|Ia |∩|Ib | , Ib ⊂ Ia |I | 1 , other wise |I |
(11)
1 1 + exp(−|μa − μb | · |σa − σb |)
(12)
I n f luenceo f bona, (b, a) = U R P(a, b) = 1 −
Sim(a, b) N W S M = Sim(a, b) J acar d · U R P(a, b)
(13)
Here, μa and μb are the mean rating of user a and b, while σa and σb represent the standard variance of user a and b, respectively.
34
A. Jindal et al.
3.8 New Heuristic Similarity Model (NHSM) This similarity metric is a variation of proximity-impact-popularity measure (PIP)[34], but NHSM comprises of three important factors—proximity, significance and similarity (PSS). Proximity considers the level of agreement between the ratings by using the absolute difference between their values. The significance factor represents how strongly an item is liked or disliked by users. Singularity represents how unique are the ratings and how it is different from the other ratings. Unlike PIP, it also considers the number of common items and user rating preference [18]. sim(u, v)PSS =
PSS(ru, p , rv, p )
(14)
p I
sim (u, v) J P SS = sim (u, v) P SS · sim(u, v) J accar d
(15)
sim(u, ν) N H S M = sim(u, ν)JPSS · sim(u, ν)U R P
(16)
4 Experiments In order to evaluate the performance of various Jaccard-based similarity measures, a series of offline experiments have been conducted. Here, we have used standardized benchmark rating datasets which are publicly available and are summarized in Table 2. Table 2 Datasets used for experimentation Dataset
Release Date Brief Description
Sparsity Level (%)
MovieLens-100 K [35] 04/1998
• Total ratings: 100,000 Over movies: 1,700 From 1,000 users
1−
100000 1000∗1700
= 94.118%
MovieLens-1 M
02/2003
• Total ratings: 1000,209 Over movies: 3900 From 6040 users
1−
1000209 6040∗3900
= 95.754%
Yahoo! Music Ratings [36]
09/2006
• Total ratings: 311,704 Over songs: 1000 From 15,400 users
1−
311704 15400∗1000
= 97.976%
Joyful Jaccard: An Analysis of Jaccard-Based …
35
4.1 Evaluation Metrics The performance of all the Jaccard-based similarity measures, which are listed in Table 1, is evaluated using two metrics—MAE and RMSE. Mean absolute error (MAE) is a measure of errors between predicted rating and actual rating. M AE =
1 | pr edicted − actual| N
(17)
Root mean square error (RMSE) is also a good measure of accuracy for recommender systems. Unlike MAE, it is more sensitive to outliers and bad predictions. RMSE =
1 ( pr edicted − actual)2 N
(18)
The lower the error values, the more efficient is the similarity measure.
4.2 Results and Discussion Figure 1a-c represents the MAE values by changing the neighborhood size for various similarity measures. These experimentally collected MAE values exhibit that the NWSM measure performed far better than all the other similarity measures under every situation. Similarly, the empirically obtained values of RMSE, shown in 2a-c, demonstrate that the NWSM measure provides the best results among all measures. All the similarity measures have performed their best in the neighborhood size of 80–120.The worst performance is shown when the neighborhood size is 20, which is too small. But after a certain threshold point, increasing the size of the neighborhood does not do any help in improving the performance of the recommender system. Generally, neighborhood size of 80–120 is found to be apt, which is neither too large nor too small (Fig. 2).
36 0.9
JMSD
MAE
0.85
CJMS D RJacc ard JacRA
0.8 0.75 0.7
20
40
60
K-Neighbours 80
100
120
140
160
180
JMSD
0.7824 0.7687 0.7595 0.7724 0.7707 0.7695 0.7695 0.7686 0.7704
CJMSD
0.8078 0.7918 0.7846 0.7835 0.7826 0.7835 0.7812 0.7799 0.7792
JacL MH Cjac MD
RJaccard 0.8135 0.7918 0.7875 0.7846 0.786 0.7848 0.7838 0.7825 0.7822 JacRA
0.7777 0.76 0.7645 0.7681 0.7733 0.7697 0.7664 0.7683 0.7689
JacLMH 0.7929 0.7757 0.7659 0.7599 0.7612 0.7595 0.7583 0.7583 0.7587 CjacMD 0.8227 0.7909 0.7911 0.7774 0.783 0.7748 0.7753 0.7772 0.7783 NWSM
0.763 0.7281 0.7279 0.7413 0.741 0.7399 0.7303 0.7301 0.7354
NHSM
0.7814 0.7652 0.7669 0.7654 0.767 0.7668 0.7696 0.7694 0.768
PCC
0.8967 0.8767 0.8739 0.8426 0.8349 0.8086 0.7926 0.7801 0.7788 (a) 1
K-Neighbours 0.95
JMS D CJM SD RJac card
0.9
JacR A
MAE
Fig. 1 MAE values for a MovieLens-100 K, b MovieLens-1 M, c Yahoo Music datasets
A. Jindal et al.
JacL MH
0.85
Cjac MD NWS M
0.8
NHS M PCC 0.75
0.7
20
40
60
80
100
120
140
160
180
JMSD
0.8033 0.7729 0.7564 0.7491 0.7432 0.741 0.7391 0.7372 0.7359
CJMSD
0.7628 0.7549 0.7514 0.7497 0.7486 0.7478 0.7472 0.7466 0.7459
RJaccard 0.7674 0.7574 0.7541 0.7522 0.7506 0.7495 0.7484 0.7477 0.7471 JacRA
0.8001 0.7718 0.7547 0.7484 0.7459 0.7399 0.7384 0.7363 0.7354
JacLMH 0.7995 0.7685 0.7514 0.7433 0.7397 0.7372 0.7346 0.7335 0.7337 CjacMD 0.8208 0.8061 0.7836 0.7695 0.7596 0.7566 0.7525 0.7495 0.7506 NWSM
0.7358 0.7292 0.7191 0.7093 0.7097 0.712 0.7122 0.7154 0.7143
NHSM
0.7858 0.7559 0.7426 0.7363 0.7331 0.7315 0.7298 0.7295 0.7286
PCC
0.9806 0.9405 0.9247 0.9193 0.9082 0.8973 0.8912 0.8822 0.8738 (b)
Joyful Jaccard: An Analysis of Jaccard-Based … Fig. 1 (continued)
37
1.15
JMS D CJM SD RJac card JacR A JacL MH Cjac MD NWS M
MAE
1.05
0.95
0.85
0.75
K Neighbours 20
40
60
80
100
120
140
160
180
200
JMSD
1.066 1.038 1.018 1.009 1.004 0.993 0.985 0.983 0.98 0.976
CJMSD
1.014 0.991 0.983 0.977 0.972 0.971 0.969 0.968 0.968 0.967
RJaccard 1.022 JacRA
1
0.99 0.984 0.979 0.977 0.975 0.972 0.971 0.97
1.073 1.04 1.015 1.004 0.995 0.987 0.983 0.981 0.976 0.973
JacLMH 1.067 1.029 1.016 1.004
1
0.997 0.993 0.987 0.983 0.981
CjacMD 1.062 1.037 1.025 1.013 1.004 1.002 0.991 0.99 0.986 0.982 NWSM
0.837 0.819 0.81 0.807 0.794 0.791 0.797 0.792 0.794 0.79
NHSM
1.07 1.028 1.009 0.998 0.993 0.983 0.977 0.975 0.971 0.969
PCC
1.125 1.094 1.08 1.074 1.065 1.061 1.059 1.055 1.052 1.047 (c)
5 Conclusion This paper explores various Jaccard-based similarity measures. Comparing all of them, experimental results show that NWSM performs well and produces better quality results as compared to other Jaccard-based similarity measures as well as traditional similarity metrics such as PCC. In the future work, more important algorithms can be tried to check the performance such as EM and sparse SVD algorithm.
38
A. Jindal et al. 1.3 JMS D CJMS D RJacc ard JacR A JacL MH Cjac MD NWS M
RMSE
1.2
1.1
1
0.9
K Neighbours 20
40
60
80
100
120
140
160
180
JMSD
1.0801 1.0148 1.006 0.992 0.9877 0.9839 0.978 0.9797 0.9796
CJMSD
1.0207 0.9942 0.9876 0.9884 0.9891 0.9853 0.9829 0.9815 0.9797
RJaccard 1.0186 0.9961 0.9867 0.9853 0.9852 0.9856 0.9852 0.9822 0.9789 JacRA
1.0713 1.0131 0.992 0.9889 0.9857 0.9804 0.9773 0.9781 0.9802
JacLMH 1.0309 1.0093 0.9993 0.9893 0.9887 0.9874 0.9853 0.9853 0.9848 CjacMD 1.0354 1.0346 1.0338 1.0214 1.003 0.9953 0.9887 0.9892 0.9875 NWSM
0.9613 0.9628 0.9935 1.0041 0.9718 0.9815 0.9921 0.9755 0.9712
NHSM
1.0505 0.9982 0.9993 0.9894 0.9873 0.9852 0.9861 0.9833 0.982
PCC
1.304 1.2523 1.1774 1.1319 1.0773 1.0605 1.0404 1.0236 1.0196
(a) 1.3
RMSE
1.2
JMS D CJM SD RJac card JacR A JacL MH Cjac MD NW SM
1.1
1
0.9
20
40
60
80
100
120
140
160
180
JMSD
1.0264 0.9911 0.9765 0.9633 0.9573 0.9523 0.9477 0.9416 0.9391
CJMSD
1.0124 0.9955 0.9877 0.9827 0.9797 0.9774 0.9757 0.9741 0.9731
RJaccard 1.0164 0.9985 0.9902 0.9848 0.9815 0.978 0.9764 0.975 0.974 JacRA
1.0279 0.9884 0.9746 0.9616 0.9541 0.9469 0.9459 0.9401 0.9389
JacLMH
1.022 0.9836 0.9725 0.9603 0.9598 0.9499 0.9445 0.9405 0.9378
CjacMD 1.0881 1.038 1.0065 0.9975 0.9864 0.9761 0.971 0.9663 0.9639 NWSM
0.9559 0.9225 0.926 0.9174 0.9103 0.9089 0.9101 0.907 0.9006
NHSM
1.0225 0.972 0.9539 0.941 0.9355 0.9308 0.9295 0.9265 0.9263
PCC
1.2777 1.2361 1.1983 1.1814 1.1623 1.1544 1.1477 1.132 1.1223
K Neighbours (b) Fig. 2 RMSE values for a MovieLens-100 K, b MovieLens-1 M, c Yahoo Music datasets
Joyful Jaccard: An Analysis of Jaccard-Based …
39
1.5
1.4 JMS D CJM SD
1.2
RJa ccar d Jac RA
1.1
JacL MH
RMSE
1.3
K Neighbours 1
20
40
60
80
100
120
Cjac MD 140
160
180
JMSD
1.4077 1.3556 1.326 1.3089 1.2965 1.2934 1.2932 1.2913 1.2848
CJMSD
1.3213 1.2958 1.281 1.2728 1.268 1.2639 1.2622 1.2596 1.2568
RJaccard 1.3285 1.2949 1.2838 1.277 1.2719 1.2676 1.2644 1.261 1.2594 JacRA
1.3975 1.3423 1.3233 1.3135 1.3026
1.3
1.29 1.2874 1.2832
JacLMH 1.4144 1.3709 1.3397 1.3223 1.3126 1.302 1.2937 1.2865 1.281 CjacMD 1.4955 1.4298 1.3932 1.3632 1.3564 1.349 1.3401 1.341 1.3286 NWSM
1.1876 1.1442 1.1199 1.083 1.0885 1.0733 1.067 1.0804 1.0827
NHSM
1.4311 1.3704 1.3379 1.3129 1.3046 1.2977 1.2901 1.2842 1.2801
PCC
1.5017 1.4379 1.4052 1.3927 1.377 1.3712 1.3641 1.3552 1.3514
(c) Fig. 2 (continued)
References 1. Ricci F, Rokach L, Shapira B, Kantor PB (2010) Recommender systems handbook, 1st edn. Springer-Verlag, Berlin, Heidelberg 2. Aggarwal CC (2016) Recommender systems: the textbook, 1st edn. Springer Publishing Company, Incorporated 3. Ekstrand MD (2011) Collaborative filtering recommender systems. Found. Trends® HumanComp Int 4(2):81–173 4. Su X, Khoshgoftaar TM (2009) A survey of collaborative filtering techniques. Adv Artif Intell Section 3:1–19 5. Breese JS, Heckerman D, Kadie C (1998) Empirical analysis of predictive algorithms for collaborative filtering. Proc 14th Conf Uncertain Artif Intell 461(8):43–52 6. Joaquin D, Naohiro I (1999) Memory-based weighted-majority prediction for recommender systems. Res Dev Inf Retr 7. Nakamura A, Abe N (1998) Collaborative filtering using weighted majority prediction algorithms. In Proceedings of the Fifteenth International Conference on Machine Learning, pp 395–403 8. Getoor L, Sahami M (1999) Using probabilistic relational models for collaborative filtering. Work Web Usage Anal User Profiling 9. Marlin B (2003) Modeling user rating profiles for collaborative filtering. In Proceedings of the 16th International Conference on Neural Information Processing Systems, pp 627–634 10. Herlocker JON, Riedl J (2002) An empirical analysis of design choices in neighborhood-based collaborative filtering algorithms. Inf Retr Boston 287–310
40
A. Jindal et al.
11. Bobadilla J, Serradilla F, Bernal J (2010) A new collaborative filtering metric that improves the behavior of recommender systems. Knowledge-Based Syst 23(6):520–528 12. Bobadilla J, Ortega F, Hernando A, Arroyo Á (2012) A balanced memory-based collaborative filtering similarity measure. Int J Intell Syst 27(10):939–946 13. Bag S, Kumar SK, Tiwari MK (2019) An efficient recommendation generation using relevant Jaccard similarity. Inf Sci (Ny) 483:53–64 14. Wu X, Huang Y, Wang S (2017) A new similarity computation method in collaborative filtering based recommendation system. In 2017 IEEE 86th Vehicular Technology Conference (VTCFall), pp 1–5 15. Lee S (2017) Improving Jaccard index for measuring similarity in collaborative filtering. Inf Sci Appl 2017:799–806 16. Suryakant, Mahara T (2016) A new similarity measure based on mean measure of divergence for collaborative filtering in sparse environment. Proc Comput Sci 89:450–456 17. Zang X et al (2017) A new weighted similarity method based on neighborhood user contributions for collaborative filtering. Proc—2016 IEEE 1st Int Conf. Data Sci Cyberspace, DSC 2016, pp 376–381 18. Liu H, Hu Z, Mian A, Tian H, Zhu X (2014) A new user similarity model to improve the accuracy of collaborative filtering. Knowledge-Based Syst 56:156–166 19. Suganeshwari G, Syed Ibrahim SP (2018) A comparison study on similarity measures in collaborative filtering algorithms for movie recommendation. Int J Pure Appl Math 119(15 Special Issue C):1495–1505 20. Sondur SD, Nayak S, Chigadani AP (2016) Similarity measures for recommender systems: a comparative study. Int J Sci Res Dev 2(3):76–80 21. Resnick P, Iacovou N, Suchak M, Bergstrom P, Riedl J (1994) Group Lens: an open architecture for collaborative filtering of netnews. In Proceedings of the 1994 ACM conference on Computer supported cooperative work, pp 175–186 22. Jaccard P (1901) Distribution comparée de la flore alpine dans quelques régions des Alpes occidentales et orientales. Bull la Socit Vaudoise des Sci Nat 37:241–272 23. Verma V, Aggarwal RK (2019) Accuracy assessment of similarity measures in collaborative recommendations using CF4J framework. Int J Mod Educ Comput Sci 11(5):41 24. Ortega F, Zhu B, Bobadilla J, Hernando A (2018) CF4J: collaborative filtering for Java. Knowledge-Based Syst 152:94–99 25. Arsan T, Koksal E, Bozkus Z (2016) Comparison of collaborative filtering algorithms with various similarity measures for movie recommendation. Int J Comput Sci Eng Appl 6(3):1–20 26. Owen S, Anil R, Dunning T, Friedman E (2011) Mahout in action. Manning Publications Co, Greenwich, CT, USA 27. Al Hassanieh L, Jaoudeh CA, Abdo JB, Demerjian J (2018) Similarity measures for collaborative filtering recommender systems. In 2018 IEEE Middle East North Africa Commun. Conf. MENACOMM, pp 1–5 28. Stephen SC, Xie H, Rai S (2017) Measures of similarity in memory-based collaborative filtering recommender system—a comparison. ACM Int Conf Proc Ser Part F1296 29. Sun SB et al (2017) Integrating triangle and jaccard similarities for recommendation. PLoS ONE 12(8):1–16 30. Verma V, Aggarwal RK (2019) A new similarity measure based on simple matching coefficient for improving the accuracy of collaborative recommendations. Int J Inf Technol Comput Sci (IJITCS) 6:37–49 31. Al-bashiri H, Abdulgabber MA, Romli A, Hujainah F (2017) Collaborative filtering similarity measures: revisiting. ACM Int Conf Proc Ser Part F1312:195–200 32. Verma V, Aggarwal RK (2020) A comparative analysis of similarity measures akin to the Jaccard index in collaborative recommendations: empirical and theoretical perspective. Soc Netw Anal Min 10(1) 33. Shardanand U, Maes P (1995) Social information filtering: algorithms for automating ‘Word of Mouth’. Proc SIGCHI Conf Hum factors Comput Syst - CHI ’95, pp 210–217
Joyful Jaccard: An Analysis of Jaccard-Based …
41
34. Ahn HJ (2008) A new similarity measure for collaborative filtering to alleviate the new user cold-starting problem. Inf Sci (Ny) 178(1):37–51 35. MovieLens | GroupLens. [Online]. https://grouplens.org/datasets/movielens/. Accessed 22 Dec 2018 36. Webscope | Yahoo Labs. [Online]. https://webscope.sandbox.yahoo.com/. Accessed 16 May 2020
An Improved Reversible Data Hiding Technique For Encrypted Images Using Difference Error Expansion Ankita Vaish, Ruchi Agarwal, and Manoj Kumar
Abstract Digital data hiding exists in many forms, one of its most useful form is reversible data hiding which recovers the secret data and extracts the cover image exactly as these were at the time of embedding. In last few years, encryption and reversible data hiding have been combined to explore a new area of research, seeking researchers’ interest in very short span of time. To enhance the payload, this paper presents a novel reversible data hiding technique for encrypted images using bilinear interpolation and difference error expansion. Difference error expansion is a technique used to embed watermark in the images, by exploiting the spatial redundancy which exists in digital images. In difference expansion schemes, correlation of adjacent pixels is exploited to create the space for embedding of secret data. This paper aims to elevate the embedding capacity and reduce the distortion effect caused by embedding with the help of difference error expansion-based reversible data hiding method. The proposed scheme investigates the use of bilinear interpolation by utilizing the pixels located at even rows and even columns for the prediction of neighbouring pixels. Notably, good results are obtained when proposed work is compared with the existing ones on the basis of embedding capacity and PSNR. Keywords Bilinear interpolation · Difference error expansion · Embedding capacity · Encryption · Reversible data hiding
1 Introduction The role of information technology in present scenario can be understood by the communication of data among people which is much easier and faster as they were A. Vaish (B) Department of Computer Science, Banaras Hindu University, Varanasi, UP, India e-mail: [email protected] R. Agarwal · M. Kumar Department of Computer Science, Babasaheb Bhimrao Ambedkar University, Lucknow, UP, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_4
43
44
A. Vaish et al.
in past. But as easily the data is available, in the same way it may be manipulated too. Forging, intruding and illegal uses of digital media are easily done by the spurious users. Thus, to protect the data from an unauthorized access, there is a need of information security. For protecting the digital content (images, audios and videos), data hiding plays a vital role. The process to hide the data into some cover media (image, audio and video) for confidential transmission of messages is referred as data hiding. Based on the property of recoverability, i.e. recovery of the original image and hidden data from the marked image, data hiding techniques can be broadly categorized into reversible and irreversible data hiding. Reversible data hiding (RDH) has been propounded to invertibly retrieve the two of embedded data and original media. In the year 1997, the first RDH algorithm was introduced by Barton [1] in the form of US patent. After that the rapid changes in technology result into extension of applications that increased the use of RDH. Literature [2–7] reports a number of techniques for RDH which can be further categorized into five classes based on difference expansion (DE) [8], histogram modification (HM) [9–12], lossless compression (LC) [13–18], prediction error (PE) [2, 19] and interpolation [20]. Recently, researchers have proposed many algorithms [21–24] in which the two main concepts, encryption and reversible data hiding, are combined to provide the certainty to both the cover image and embedded data. Reversible data hiding in encrypted images (RDHEI) are widely used in perceptive areas, namely artworks, defence, law forensics, medical imaginary and many more, here slightest distortion is forbidden, and certainty of original media is as significant as the secret data. The basic idea behind introducing RDHEI is to hide the original image content from unauthorized users as well as data hider (if data hider and content owner are different). RDHEI has been further divided into two groups depending on the time of encryption and creation of embedding room as follows: (1) vacating room before encryption (VRBE) [22, 25, 26] and (2) vacating room after encryption (VRAE) [21, 24, 27–29]. In VRBE methods, the space for embedding is created before encryption, i.e in plaintext domain. Thus, VRBE methods require extra pre-processing for reserving space before encrypting the image. First VRBE method has been proposed by Ma et al. [22] by making use of LSBs of some selected pixels into others by using existing RDH methods for creating embedding room, and then encryption followed by data hiding is performed. Zhang et al. [26] presented an improved RDH method by applying prediction error technique (PE), the predicted errors are encrypted by encryption schemes, and further data embedding has been done by histogram shifting of calculated predicted errors. In [30], a patch-level sparse representation scheme has been utilized to make more space for embedding secret data. Shiu et al. [25] introduced a DE-based RDH method using Paillier homomorphic encryption [31] to reserve a space for embedding data. Agarwal et al. [32] proposed a new scheme based on mean value with a limited embedded capacity. RDHEI using interpolation technique has been introduced by Malik et al.[33] having payload about 12 bit per pixel (bpp). In VRBE methods, embedding rate up to 0.5 bpp can be attained, but it might
An Improved Reversible Data Hiding Technique …
45
be ineffectual due to requirement of extra pre-processing of images. Overcoming this issue, VRAE methods [21, 24] are proven more efficient. In VRAE algorithms, the owner first encrypts the original image, and then the data hider embeds secret data in the encrypted image. The first technique has been designed by Zhang [24] in which encryption of image is performed by XORing a random number sequence bitwise with the original image pixel values, the image is grouped into the blocks, and then secret data is inserted (one bit per block) flipping the three LSBs in each block. This method cannot remove error during data extraction because of its block-dependent constraint. Hong et al. [21] put forward a refined variant of Zhang’s scheme [24] with the introduction of function calculating smoothness of blocks and additionally using side match for correctly and incorrectly retrieved blocks. Although his approach gave better results compared to Zhang’s method, complete reversibility was not attained. In [27], an improved smoothness function measuring block’s complexity was developed. Unlike Hong’s scheme, it covered border pixels. A RDHEI utilizing shifting of histogram with an average embedding rate of 0.5 bpp was proposed by Zhang et al. [29]. Yi et al. [28] introduced a prediction error expansion approach-based RDHEI by making use of block-level prediction error expansion. All of the schemes discussed above share some common disadvantages like limited payload capacity and low PSNR. The major challenges to augment the working of prediction error-based RDH are as follows: maximization of accurate prediction, minimization of embedding distortion, and side information. Based on these challenges and with the aim to increase the embedding capacity as well as visual quality, we propose an improved RDH technique for encrypted images using prediction error expansion. This approach predicts the values of all the pixels other than at even row and even column positions by use of bilinear interpolation [34]. To handle the problem of overflow and underflow, a location map is created as a side information. By using this methodology, the embedding capacity is increased to three-fourth times of original image, which was earlier approximately half in the existing scheme [33]. Since embedding is not done at pixels located at even rows and even columns positions which results into less distortion, effectiveness of the scheme can be seen by high embedding capacity and PSNR values among original and directly decrypted images (containing secret data). The rest of this paper is organized as follows: Sect. 2 explains the fundamental concept of prediction error expansion, the proposed scheme is discussed in Sect. 3, analysis of experimental results is presented in Sect. 4, and in Sect. 5 conclusions are derived from the proposed work.
2 Prediction Error Expansion The basic concept of prediction error expansion-based techniques is to first predict the pixels of cover image using adjacent pixels, and then the difference between
46
A. Vaish et al.
original and predicted pixel value is calculated. These difference values are called predicted errors further used for embedding the secret data when expanded. We briefly explain the Kumar’s method [35] which utilizes the difference expansion idea [8] in this section. Let ai, j and aˆ i,2 j be the original pixel values and the predicted values at even columns positions of input image, respectively. The predicted values are calculated by taking mean of neighbouring pixel values of odd a +a columns (except last column) as follows: aˆ i,2 j = i,2 j−1 2 i,2 j+1 . The prediction errors are obtained as follows: pi,2 j = ai,2 j − aˆ i,2 j , where ai,2 j is the original pixel value positioned at even columns. The secret bit embedding is done by expanding prediction errors at even columns. Let m k be the k th bit of secret data and p i,2 j be the modified predicted error as follows: p i,2 j = 2 × pi,2 j + m k ; where m k = 1 or 0. Calculation of the embedded pixel values is done by performing addition of predicted and modified predicted error values as follows: aˆ i,2 j = aˆ i,2 j + p i,2 j . The range of embedded pixel values aˆ i,2 j lies between [0, 255]. If the value of embedded pixels is below 0 or above 255, i.e. aˆ i,2 j < 0 or xˆ i,2 j > 255, they are called as underflow and overflow values, and location map (LM) is created as in [8] for such values by marking 1 at respective places. The underflow and overflow values are replaced with original ones. LM is compressed using arithmetic coding and embedded into pixels positioned at odd columns, and in this way, the final embedded image I has been obtained. Extraction procedure is followed by finding the positions of overflow and underflows pixels by uncompressing location map from odd columns. Recovery of original image is performed by predicting pixel value aˆ i,2 j again as aˆˆ i, j , and prediction error is obtained as p i,2 j = aˆ i,2 j − aˆˆ i,2 j . The extraction of secret bits is performed as follows: m i = p i,2 j mod 2, original predicted error was calculated p
by p i,2 j = 2i,2 j , where the function a rounds the a to the closest integer value pointing to the minus infinity. The original pixels are performed as follows: a i,2 j = aˆˆ i,2 j + p i,2 j . For complete reversibility, the predicted value must be recovered during extraction. We propose the method for encrypted images which utilizes the basic idea of predicting the values which can be further recovered during the process of extraction. The proposed scheme utilizes the pixels at even rows and even columns for prediction of all other pixel values by bilinear interpolation which increases the embedding capacity by a factor 43 which was earlier approximately 21 in existing scheme [33]. Thus, in proposed work the embedding capacity reaches 43 of original image. The methodology of proposed scheme is explained in the following section.
3 Proposed Scheme The generalization of the proposed methodology is depicted in Fig. 1. The figure incorporates three parties, namely content owner, data hider and receiver. The role of content owner is to encrypt the original image by generating a pseudorandom matrix
An Improved Reversible Data Hiding Technique …
47
Fig. 1 Generalization of proposed work
(similar in size with original image) with the help of seed value, and then addition modulo 256 of random sequence with original image is performed. After encryption, concept of difference error expansion is utilized , and interpolated image is obtained using bilinear interpolation by data hider. The prediction errors are calculated, and the expansion of prediction error is done for embedding the secret data. Now this marked encrypted version of original cover is communicated with the help of secure channel at the receiver’s end. Receiver then extracts and recovers the secret data and original image using data hiding and encryption keys, respectively. The detail description of proposed scheme is explained in following subsections:
3.1 Encryption Phase The encryption technique uses additive modulo 256. As additive modulo encryption has property that the computation time on original and encrypted covers is same, for fast computation time additive modulo 256 encryption scheme is used. Encryption of original image I consists of two parts: 1. Generate a pseudorandom number matrix R with the help of pseudorandom number generator and secret seed value. 2. Take addition modulo 256 of pseudorandom number matrix R with original image I to obtain encrypted image E. E i, j = (Ii, j + Ri, j ) mod 256
(1)
where i and j lie between [1, r ] and [1, c], respectively, with r × c as image size. Figure 2 shows the six standard test images with their encrypted versions obtained after applying above encryption scheme. We use pseudorandom number generator to generate the pseudorandom matrix as its addition modulo with original image will also give pseudorandom number matrix, and the research till date says there exists no probability polynomial time scheme which can differentiate between random numbers and pseudorandom numbers matrix [36]. Use of homomorphic encryption saves computation time, and also the addition modulo encryption is robust against any probability polynomial time adversary.
48
A. Vaish et al.
Fig. 2 Standard test images a Lena, b Baboon, c Boat, d Peppers, e Aeroplane, f Sail and their respective encrypted images
3.2 Data Embedding Phase The proposed work utilizes the basic fundamentals of traditional difference error expansion in encrypted images, the predicted errors are calculated, and the expanded prediction errors are used for embedding the additional or secret data. In proposed scheme, the secret bits are embedded into the pixels excluding those located at even row and even column positions (2i, 2 j). To attain this, the encrypted image is divided into four subimages based on the position of pixels as shown in Fig. 3: & indicates pixel positioned at odd row and odd column (OO), * indicates pixel positioned at odd row and even column positions (OE), # indicates pixel positioned at even row and odd column (EO), and $ indicates pixel positioned at even row and even column
Fig. 3 Decomposition of image into four subimages
An Improved Reversible Data Hiding Technique …
49
Fig. 4 a Embedding procedure, b Extraction procedure
(EE), as shown in Fig. 3. The subimage EE is used to predict the values of other subimages OO, OE and EO. Prediction error is calculated by subtracting interpolated image from encrypted image. The pixel values of EE subimage are kept reserved for obtaining the original predicted values and prediction errors of all other subimages. To overcome the condition of underflow and overflow, location map is created in the form of the side information. The detailed steps of embedding algorithm are explained in Fig. 4a. During embedding, the pixel values other than pixels located at (2i, 2 j) are predicted by applying bilinear interpolation on subimage EE. The next step is to calculate the prediction error by using the predicted and original pixel value. Expansion of prediction error is performed for embedding the secret bits, and modified predicted errors are calculated
50
A. Vaish et al.
by adding watermark bits and expanded prediction errors. The marked pixel values are calculated by adding modified prediction errors and predicted values. If marked pixel values are below 0 or above 255, formation of location map takes place by putting 1 at these places. The pixels at underflow or overflow positions are replaced with the original ones. Compression of location map is performed by using wellknown arithmetic coding, and the compressed form for the same is embedded in the subimage EE. After this, the final embedded encrypted image is obtained. The steps of data embedding are summarized in the following algorithm. Embedding Algorithm 1. Let E be the original encrypted image and divide it into four subimages (as shown in Fig. 3) according to row and column positions as follows: EE, OE, EO and OO. 2. Generate the interpolated image Eˆ by predicting the pixel values from the encrypted image using bilinear interpolation for all the pixels except subimage EE as EE is used to interpolate the other subimages. 3. Calculate difference error P by following operation: Pi, j = E i, j − Eˆ i, j
(2)
where E i, j and Eˆ i, j are the original and interpolated pixel values of encrypted image at the (i, j)th position. 4. Expand the difference errors of pixels except subimage EE for embedding the secret bits. Let wk be the k th bit of secret data, calculate the modified difference error P i, j as follows: (3) P i, j = 2 × Pi, j + wk ; given, wk = 0 or 1 5. Calculate marked pixel values E i, j by adding interpolated pixel values and modified prediction error values E i, j = Eˆ i, j + P i, j
(4)
6. Location map L is created at this point for the locations having value below 0 or above 255, and the overflow and underflow values of marked image E i, j are replaced with the pixel values of encrypted image E i, j . 7. For reducing the length of location map, it is compressed with the help of arithmetic coding, and compressed form is embedded into subimage EE. After all these operations, the final marked encrypted image E i, j is obtained.
An Improved Reversible Data Hiding Technique …
51
3.3 Extraction Phase The extraction procedure is carried out by a valid receiver, and during this phase, extraction and decryption operations are performed on the basis of availability of data hiding and decryption keys. If both the keys are available, decryption is followed by extraction of secret data, and the obtained result (secret data and cover image) is exactly identical to the original data. On the availability of only decryption key, the image obtained is similar to original version (notably good PSNR). The extraction procedure is explained in Fig. 4b. During extraction, marked image is divided into four subimages similar as in embedding procedure. Location map is uncompressed for finding the positions of watermarked pixels which suffer from overflow and underflow conditions. Apply extraction procedure on all the subimages except subimage EE, and predict the values for pixels using subimage EE. The steps of extraction are summarized in the following algorithm. Extraction Algorithm 1. Let the marked encrypted image be E i, j , decompose it in four subimages similar as in embedding. 2. Uncompress the location map L to obtain the overflow and underflow positions from subimage EE. 3. Predict the values of interpolated image Eˆ i, j for all the pixels except subimage EE, using bilinear interpolation on subimage EE of watermarked image. 4. Obtain difference error Pˆi, j for all the pixels except subimage EE, overflow and underflow positions as follows: Pˆi, j = E i, j − Eˆ i, j
(5)
5. Extract secret bits wi as follows: wi = Pˆi, j mod 2
(6)
6. Calculate original difference error P i, j as follows: P i, j =
Pˆi, j 2
(7)
7. Recover the encrypted image E i, j by the following formula: E i, j = Eˆ i, j + P i, j
(8)
8. At the locations where embedding is not performed, put the pixel values from subimage EE, overflow and underflow positions into encrypted image without any modification.
52
A. Vaish et al.
Image Recovery Recovery of original image is done by performing reverse operation of encryption phase. After extraction of secret bits with the help of encryption key, the valid receiver will generate the same pseudorandom number matrix R. By using following formula, the original image is recovered as follows: Iˆi, j = (E i, j − Ri, j )mod256
(9)
where Iˆi, j is the recovered image, E i, j is encrypted image after extraction of secret bits, and Ri, j is the pseudorandom number matrix generated with the help of encryption key. The recovered image and secret bits are identical to the original image and hidden data bit by bit and thus free from any error and attains full reversibility.
4 Experimental Analysis The proposed work is verified against many experiments, implemented in MATLAB with different standard test images (Fig. 2) each having 512 × 512 as size with pixel values ranging from [0 − 255]. Figure 5e shows the decrypted image after data extraction, and Fig. 5f is the absolute difference between original Lena image (Fig. 5a) and recovered image (Fig. 5e) having all pixels values as 0, which shows exact recovery of original image. Image quality, embedding rate and security are some of the factors which validate any RDHEI scheme. Following are some parameters on which feasibility of proposed scheme is tested.
4.1
Based on Payload Limit
The proposed algorithm is tested using six standard test images (shown in Fig. 2). The pixel values positioned at EE have been utilized for predicting the pixel values positioned at other locations by using bilinear interpolation. Thus, for any digital image with size r × c, the maximum payload limit would be 43 × (r × c) bits, for instance, if image size is equal to 512 × 512, the result of maximum payload would be 196608 bits. But the actual payload capacity would be less than 196608 as it
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 5 a Lena image, b Encrypted version of (a), c Embedded encrypted image, d Directly decrypted image, e Recovered image, f Difference between (a) and (e)
An Improved Reversible Data Hiding Technique …
53
depends on number of pixels having overflow or underflow. Suppose the number of pixels having underflow or overflow is b , then the actual embedding capacity would be (196608-b) bits. Here in this scheme, the compressed form of location map is embedded into subimage EE. Thus, embedding of location map would not decrease the embedding rate as the embedding of secret bits is done in the subimages other than EE. Figure 5d justifies the statement as it is identical to original image (Fig. 5a). Therefore, the pure payload limit for any image with size r × c would be ( 43 × (r × c) -b) bits.
4.2 Based on Security Statistical Parameters Table 1 shows the resultant values obtained between original and marked encrypted image on applying different statistical parameters discussed below. 1. Correlation coefficient: Spatial bonding between the input and marked images is calculated with the help of correlation coefficient, it tells how strongly the two images are correlated, and its value lies between –1 and 1. Table 1 clearly shows that the values between marked encrypted and original images are nearly zero. Thus, apparently it can be said that the two images are almost different or no correlation exists between them, and hence, no information about original cover from the marked encrypted image can be leaked out. 2. Entropy: Measure of randomness or average uncertainty in an image is interpreted as entropy. Table 1 shows that the results of entropy for all the images are nearly equal to 8; therefore, marked encrypted medium can interpret nothing about original cover. 3. PSNR: The perceptible nature of an image is defined by peak signal-to-noise ratio (PSNR). The PSNR vaue of an image with reference to other specifies the visual resemblance of first with the reference image. Thus, high PSNR means high resemblance and vice versa. Table 1 displays the result of PSNR values
Table 1 Values of marked encrypted images with original images based on the discussed parameters using proposed scheme Images Correlation PSNR SSIM Entropy coefficient Lena Baboon Boat Peppers Aeroplane Sail
0.0009 0.0013 0.0005 0.0026 0.0027 –0.0017
7.9668 8.4889 8.6127 8.2577 7.5355 7.6861
0.0067 0.0077 0.0082 0.0078 0.0085 0.0072
7.8572 7.8779 7.7758 7.9349 7.3739 7.7183
54
A. Vaish et al.
between marked encrypted and original image which are extremely low for any visual resemblance representing the effectiveness of encryption scheme; hence, recovery of original data without knowledge of keys is not possible. 4. SSIM: Measure of image quality degradation is termed as structural similarity index metric (SSIM). Range of SSIM is between –1 and 1. Table 1 shows that all SSIM values are near to zero which means that original and marked encrypted images are totally different. All these statistical parameters measured between original and marked encrypted images show the randomness and unpredictable nature of marked encrypted image which secures our scheme against any statistical attack. Based on Histogram Distribution The histogram analysis of original and encrypted versions of images is depicted in Figs. 6 and 7, respectively. As histogram distribution is used to see the counts of pixels based on their respective intensities values by which one can get the actual information about an image, the uniform distribution of histogram of encrypted images shows the unpredictability of original values. From Fig. 7, it can be explicitly said that attackers would not be able to get any information about original values, and hence, the encryption scheme proves to be secure against attacks. 4000
3000 3000
2500
3000
2500
2000
2000
1500
2000
1500
1000
1000
500
500
0
1000
0 0
50
100
150
200
250
0 0
50
100
150
200
250
0
50
100
(b)
(a)
150
200
250
150
200
250
(c)
3000
3000
2500
4000
2000
2500
3000
2000
1500
1500
2000
1000
1000 1000
500
500
0
0 0
50
100
150
200
250
0 0
(d)
Fig. 6 Histogram of original images
50
100
150
(e)
200
250
0
50
100
(f)
An Improved Reversible Data Hiding Technique …
55
2500
2500
2500
2000
2000
2000
1500
1500
1500
1000
1000
1000
500
500
500
0
0 0
50
100
150
200
0 0
250
50
100
(a)
150
200
250
0
2500
2500
2000
2000
2000
1500
1500
1500
1000
1000
1000
500
500
500
0
0 50
100
100
150
200
250
150
200
250
150
200
250
(c)
2500
0
50
(b)
0 0
(d)
50
100
150
(e)
200
250
0
50
100
(f)
Fig. 7 Histogram of encrypted images
4.3 Comparison Comparison of the proposed method has been done with recently published reversible data hiding techniques for encrypted images [33]. Effectiveness of the proposed method can be seen in Table 2 which describes that the proposed method provides better PSNR of directly decrypted image in comparison with interpolation-based scheme [33]. Table 2 compares the existing [33] and proposed scheme on the basis of PSNR values for several test images at same payload. Observation from Table 2 is elaborated as follows: 1. the proposed scheme achieves higher PSNR on comparing with the existing scheme [32, 33], depicting the efficacy of the proposed method. Nearly, all the standard test images on which the proposed algorithm has been applied, and it gives optimistic results on comparing the PSNR values with the existing scheme [33]. For instance, for the payload of 0.48 bits per pixel, the PSNR value of directly decrypted Lena image is 16.427db in proposed scheme which is 12.056db in the existing scheme [33] for the same directly decrypted Lena image at the same embedding capacity. Results obtained after comparing proposed scheme with [33] are shown in Table 2. Proposed scheme provides better PSNR at the same embedding in comparison with existing schemes [33]. Therefore, the proposed scheme proves more applicable than existing scheme [33] in terms of both the factors payload and perceptibility.
56
A. Vaish et al.
Table 2 PSNR-based comparison between proposed scheme and [33] for directly decrypted images Test images Payload (bpp) [33] Proposed scheme Lena Baboon Boat Peppers Aeroplane Sail
0.4833 0.4641 0.4816 0.4820 0.4829 0.4832
12.056 12.042 12.037 12.037 12.043 12.045
16.427 15.959 16.178 16.213 15.603 16.116
5 Conclusion An improved RDHEI scheme based on difference error expansion is presented in this paper. The correlation among the neighbouring pixel is less in encrypted images, but it has been nicely utilized in the proposed work by using homomorphic additive modulo 256 encryption scheme as a subimage EE has been used to predict the pixel value of other subimages. Due to bilinear interpolation, most of the values in the prediction error are almost close to zero, which results in good visual quality even after embedding. The maximum payload using the proposed method for any image with size r × c is (( 43 × (r × c)) - l) bits, where l is the number of pixels having underflow or overflow. The payload limit can be further raised by exploring the level of embedding in the proposed scheme, but it may cause increase in the computational time. Comparison of proposed work with novel RDHEI techniques has shown higher PSNR and embedding rate. Thus, potency of the proposed work can be seen on the ground of payload with good image quality and security.
References 1. Barton JM (1997) Method and apparatus for embedding authentication information within digital data. US Patent 5,646,997 2. Hong W, Chen TS, Shiu CW (2009) Reversible data hiding for high quality images using modification of prediction errors. J Syst Softw 82(11):1833–1842 3. Ni Z, Shi YQ, Ansari N, Su W (2006) Reversible data hiding. IEEE Transac Circ Syst Video Technol 16(3):354–362 4. Thodi DM, Rodrguez JJ (2007) Expansion embedding techniques for reversible watermarking. IEEE Transact Image Proc 16(3):721–730 5. Van Leest A, Veen M, Bruekers F (2003) Reversible image watermarking. ICIP 2003 Proc 2003 Intern Conf Image Proc 3:I–731–734. IEEE 6. De Vleeschouwer C, Delaigle JF, Macq B (2001) Circular interpretation of histogram for reversible watermarking. In Proceedings of the IEEE 4th Workshop on Multimedia Signal Processing, pp 345–350 7. Xuan G, Yang C, Zhen Y, Shi YQ, Ni Z (2004) Reversible data hiding based on wavelet spread spectrum. In Multimedia Signal Processing, 2004 IEEE 6th Workshop on, pp. 211–214. IEEE
An Improved Reversible Data Hiding Technique …
57
8. Tian J (2003) Reversible data embedding using a difference expansion. IEEE Trans Circuits Syst Video Techn 13(8):890–896 9. Agrawal S, Kumar M (2015) An improved reversible data hiding technique based on histogram bin shifting. Smart Innov Syst Technol 43(6):239–248 10. Li X, Li B, Yang B, Zeng T (2013) General framework to histogram-shifting-based reversible data hiding. IEEE Trans Image Proc 22(6):2181–2191 11. Li X, Zhang W, Gui X, Yang B (2013) A novel reversible data hiding scheme based on twodimensional difference-histogram modification. IEEE Trans Inf Foren Sec 8(7):1091–1100 12. Wang ZH, Lee CF, Chang CY (2013) Histogram-shifting-imitated reversible data hiding. J Syst Soft 86(2):315–323 13. Fridrich J, Goljan M, Chen Q, Pathak V (2004) Lossless data embedding with file size preservation. In Electronic imaging, pp 354–365. International Society for Optics and Photonics 14. Fridrich J, Goljan M, Du R (2001) Invertible authentication. In Photonics West 2001-Electronic Imaging, pp 197–208. International Society for Optics and Photonics 15. Fridrich J, Goljan M, Du R (2002) Lossless data embedding for all image formats. In Electronic Imaging, pp 572–583. International Society for Optics and Photonics 16. Fridrich J, Goljan M, Du R (2002) Lossless data embedding: new paradigm in digital watermarking. EURASIP J Appl Signal Proc 2002(1):185–196 17. Guo J, Liu Z, Liu S (2007) Watermarking based on discrete fractional random transform. Optics Commun 272(2):344–348 18. Liu Z, Xu L, Guo Q, Lin C, Liu S (2010) Image watermarking by using phase retrieval algorithm in gyrator transform domain. Optics Commun 283(24):4923–4927 19. Feng JB, Lin IC, Tsai CS, Chu YP (2006) Reversible watermarking: current status and key issues. IJ Network Sec 2(3):161–170 20. Govind PV, Wilscy M (2015) A new reversible data hiding scheme with improved capacity based on directional interpolation and difference expansion. In International Conference on Information and Communication Technologies, ICICT 2014, 3–5 December 2014 21. Hong W, Chen T, Wu H (2012) An improved reversible data hiding in encrypted images using side match. IEEE Signal Process Lett 19(4):199–202 22. Ma K, Zhang W, Zhao X, Yu N, Li F (2013) Reversible data hiding in encrypted images by reserving room before encryption. IEEE Trans Inf Forens Sec 8(3):553–562 23. Shi Y, Li X, Zhang X, Wu H, Ma B (2016) Reversible data hiding: advances in the past two decades 24. Zhang X (2011) Reversible data hiding in encrypted images. IEEE Signal Process Lett 18(4):255–258 25. Shiu Chih-Wei, Chen Yu-Chi, Hong Wien (2015) Encrypted image-based reversible data hiding with public key cryptography from difference expansion. Image Commun 39:226–233 26. Zhang W, Ma K, Yu N (2014) Reversibility improved data hiding in encrypted images. Signal Proc 94(1):118–127 27. Liao X, Shu C (2015) Reversible data hiding in encrypted images based on absolute mean difference of multiple neighboring pixels. J Visual Commun Image Repres 28:21–27 28. Yi S, Zhou Y (2017) Binary-block embedding for reversible data hiding in encrypted images. Signal Proc 133:40–51 29. Zhang R, Lu C, Liu J (2019) A high capacity reversible data hiding scheme for encrypted covers based on histogram shifting. J Inform Sec Appl 47:199–207 30. Cao X, Du L, Wei X, Meng D, Guo X (2016) High capacity reversible data hiding in encrypted images by patch-level sparse representation. IEEE Trans Cybern 46(5):1132–1143 31. Paillier P (1999) Public-key cryptosystems based on composite degree residuosity classes. In Proceedings of the international conference on the theory and applications of cryptographic techniques, pp 223–238. Springer 32. Agarwal S, Kumar M (2017) Mean value based reversible data hiding in encrypted images. Optik 130:922–934 33. Malik A, Wang H, Chen T, Yang T, Khan NA (2019) Reversible data hiding in homomorphically encrypted image using interpolation technique. J Inform Sec Applic 48
58
A. Vaish et al.
34. Mastylo M (2013) Bilinear interpolation theorems and applications. J Funct Anal 265(2):185– 207 35. Kumar M, Agrawal S (2016) Reversible data hiding based on prediction error expansion using adjacent pixels. Sec Commun Netw 9(16):3703–3712 36. Xinpeng Z, Guorui F, Yanli R, Zhenxing Q (2012) Scalable coding of encrypted images. IEEE transactions on image processing: a publication of the IEEE Signal Processing Society 21:3108–3114
Assessment of Image Quality Metrics by Means of Various Preprocessing Filters for Lung CT Scan Images Sugandha Saxena, S. N. Prasad, and T. S. Deepthi Murthy
Abstract Lung cancer is the most predominant and lethal growth around the world. Lung Computerized Tomography (CT) scan images have been contributing enormously to clinical research and diagnosis. Although due to the presence of artifacts, noise and blurring effects, the CT images produce a degraded output of the actual part under diagnosis. Hence, application of preprocessing filters on CT images proves crucial to reduce the noise and improve the quality of image. In this paper, we evaluate the image quality by exploring various parameters such as Mean Square Error (MSE), Peak Signal to Noise Ratio (PSNR), Speckle reduction and Mean Preservation Index (SMPI) and Speckle Suppression Index (SSI), thereby measuring the performance of different filters. Keywords Lung CT scan images · Lung cancer · Preprocessing filter · Image quality metric
1 Introduction Lung cancer or lung carcinoma [14] is an uncontrollable growth of epithelial cells blocking the respiratory tract and is a major cause of death globally. Early detection increases the survival rate of a patient from 15 to 50% [5]. It also has been an enormous healthcare problem across the world [6]. Lung cancer can be classified as small cell cancer and non-small cell cancer. Most of the people get affected by non-small cell cancer. Among the conventional ways for detection of lung cancer such as X-ray, CT scan, Magnetic Resonance Imaging (MRI) scan, Positron Emission Tomography (PET) scan, CT scan is preferred as it creates detailed cross sectional images of the body parts under diagnosis through the use of X-rays and is capable of detecting a S. Saxena (B) · S. N. Prasad · T. S. D. Murthy School of Electronics and Communication Engineering, REVA University, Bangalore, Karnataka, India T. S. D. Murthy e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_5
59
60
S. Saxena et al.
very small nodule in the lung [1]. On the downside, CT scan images are extremely prone to noise which reduces the image clarity, thereby, reducing accuracy of the radiologist [2, 8]. In order to reduce the risk of inaccurate diagnosis and analysis, image de-noising is a major task in image processing. There are many filtering techniques to improve the image visualization and ease out diagnostic process. The objective of this paper is to apply preprocessing filters to enhance image quality by eliminating unwanted noise while keeping intact the input image details.
2 Methodology In medical applications, accuracy is utmost important because we can not put human lives into risk. In this section, performance of preprocessing filters is measured by adopting the following methodology: I. II. III. IV.
Acquisition of lung CT scan image. Application of different filtering techniques on lung CT scan images. Evaluating various image quality parameters. Analyzing to identify the effective filtering technique.
In this work, filtering techniques applied are median filter, average filter, Wiener filter, Laplacian filter and Gaussian filter. These techniques will eradicate the unwanted noise and enhance the image quality which is known as preprocessing stage. Using MATLAB software all the filtering techniques has been executed. The flow chart of methodology is shown in Fig. 1. Fig. 1 Flow chart of the methodology
Input CT Scan Image
Apply Different Filtering Techniques
Calculate Image Quality Metric by Different Filters
Recognize the Effective Filter Which Eliminates Noise
Assessment of Image Quality Metric …
61
2.1 Input CT Scan Image Acquisition of lung CT scan images is the first step. During this process, many factors can introduce noise to the image. Depending on the type of the noise, filtering techniques can be applied [9]. In this paper, twenty lung CT scan images are taken from cancer imaging archive repository. Initially, all the images were downloaded in the DICOM format. For further preprocessing, images were first converted in JPEG format and then into gray scale images. After this, input images are ready for preprocessing techniques.
2.2 Filtering Techniques Several filtering techniques are used in order to eliminate the noise while restoring the images.
2.2.1
Median Filter
One of the nonlinear method is median filter which is capable of removing different types of noise to a greater extent and restores the sharpness of the image [4]. It uses 3X3 mask which is moved over the entire image to replace all the elements of the image by its median value [3]. The median can be calculated by placing the mask on the first element of the image. All the elements of the mask are then sorted in ascending order, and median is calculated [17]. The mask is then slided to the next element of the image until all the elements of the image are covered.
2.2.2
Average Filter
Average filter is another way of eliminating spatial noise from the input image. In this technique, input image matrix “X” with V rows and U columns will be considered. Then a matrix of V + 2 rows and U + 2 columns should be built by adding zeros to the input image matrix. After this, mask of 3X3 will be kept on the first element of the input image matrix X and average of all the elements present in the mask will be calculated. This average value replaces the element X(1,1) of the input image matrix X. By shifting the mask to the next succeeding element of the input matrix and repeat the process by replacing all the input matrix elements by the average value of its neighboring pixels [17].
62
2.2.3
S. Saxena et al.
Wiener Filter
Wiener filter is a linear method which offers an amazing tradeoff among reverse filtering and noise smoothing. Using statistical approach, additive noise can be removed by using Wiener filter. It is ideal for calculating MSE, local mean and variance around each pixel. In Fourier domain, Wiener filter can be expressed by the following Eq. 1 [19, 20]: S(f1,f2)=
H*(f1,f2)Syy(f1,f2) H(f1,f2)2Syy(f1,f2)+Snn(f1,f2)
(1)
where. Syy (f1,f2)—power spectra of original image. Snn (f1,f2)—power spectra of additive noise. H (f1,f2)— blurring factor of filter.
2.2.4
Gaussian Filter
Gaussian filter uses 2D distribution and preserves the edges of the input image. This convolves the 2D Gaussian distribution function with the image [12]. This filter works in two stages when applied to an image. First stage is filtering of each pixel in the input image in the horizontal direction by centering the filter on that pixel values. Later, all the pixel values are multiplied by the weight at each filter location to get the resulting new pixel values. Second stage is filtering all the pixels of horizontal processed image in vertical direction to get the final image. The Gaussian G(x,y) of an image can be calculated as per the Eq. 2 given below: G(x, y) =
2.2.5
−x 2 + y 2 1 exp 2 2π σ 2σ 2
(2)
Laplacian Filter
Laplacian filter is an edge detector and can be utilized to detect the edges of an input mage. This filter calculates the second derivative of an image by measuring the rate of change in first derivative of an input image. Further, kernel of weights will be convolved with each pixel value and its neighboring pixel in an input image in order to get the new pixel values. It is also used to enhance feature with sharp discontinuity [12]. The Laplacian L(x,y) of an image can be calculated as per the Eq. 3 given below: L(x, y) =
∂2 I ∂2 I + ∂x2 ∂ y2
(3)
Assessment of Image Quality Metric …
63
2.3 Performance Parameter Measure The process of image acquisition degrades the image quality due to noise, blurring and artifacts. The image quality has been measured in terms of Mean Square Error (MSE), Peak to Signal Noise Ratio (PSNR), Speckle Suppression and Mean Preservation Index (SMPI) and Speckle Suppression Index (SSI) after the application of different filters on lung CT scan images. Their measuring parameters have been compared and analyzed in order to identify the filter technique which can reduce noise from an image effectively.
2.3.1
Mean Square Error (MSE)
Visible distortions in image can be measured by MSE. It can be defined as the difference between pixel values of estimated and true image [13]. An image with high MSE value shows more distortion compared to an image with low MSE values [10]. It can be calculated by using Eq. 4 which describes two monochrome images (X,Y) each of axb size where one image is noisy approximation of the other image. 2 1 X(i,j)-Y(i,j) ab i=0 j=0 a-1 b-1
MSE =
2.3.2
(4)
Peak Signal to Noise Ratio (PSNR)
It is a quality measurement between original and noisy image. It is a ratio of maximum pixel value (Maxi) to the MSE value of the image [13]. MSE value is calculated beforehand in order to estimate PSNR value. High-quality image always indicates high value of PSNR and can be calculated by using Eq. 5: PSNR=10log10 [
2.3.3
Max2i ] MSE
(5)
Speckle Suppression Index (SSI)
It measures the performance of filter. The value of index should be less than 1 which indicates that the filter can effectively remove the speckle noise [16]. It can be calculated by using Eq. 6, where If and Io indicate filtered and original image, respectively.
64
S. Saxena et al.
√ Var(If )*mean(Io ) SSI = √ Var(Io )*mean(If )
2.3.4
(6)
Speckle Suppression and Mean Preservation Index (SMPI)
SMPI is calculated to assess the effectiveness of filters. Lower value of this index indicates high performance of filter [16]. SMPI can be calculated by using Eq. 7, where Io and If are original and filtered images, respectively.
V ar I f SM P I = Q ∗ √ V ar (Io )
(7)
where Q = 1 + |mean(Io ) − mean I f |
2.4 Results and Discussion The experimentation has been done by applying Gaussian, Laplacian, wiener, average and median filter on 20 lung CT scan images. Image quality parameters such as mean square error (MSE) have been calculated and plotted in graph as shown in Table1 and Fig. 2. Peak Signal to Noise Ratio (PSNR) has been calculated and plotted in graph as shown in Table 2 and Fig. 3. Speckle Suppression Index (SSI) has been calculated and plotted in graph as shown in Table 3 and Fig. 4. From the results, it is observed that SSI values for Laplacian filter is greater than 1 which indicates bad performance. Hence, Laplacian filter can be neglected while plotting the graph. Speckle Suppression and Mean Preservation Index (SMPI) have been calculated and plotted in graph as shown in Table 4 and Fig. 5. From the results, it is observed that SMPI values for Laplacian filter are very high compared to other filters and indicate poor performance. Hence, Laplacian filter can be neglected while plotting the graph. For a high-quality image, the MSE value should be low, PSNR value should be high, SSI value should always be less than 1 and SMPI should be as low as possible [10]. From the above results, it is clearly seen that Laplacian filter is best suited for preprocessing stage but at the same time, it has highest SMPI and SSI values which indicates that it is not effective in removing speckle noise. Therefore, we consider median filter over Laplacian filter which shows low MSE value, high PSNR
Assessment of Image Quality Metric …
65
Table 1 MSE values calculated for 20 sample images Sample images Gaussian filter Laplacian filter Wiener filter Average filter Median filter Image 1
4.2167
1.6770
4.8635
6.3060
2.5365
Image 2
4.8325
1.8847
6.4028
7.4416
3.5216
Image 3
4.9403
1.8681
6.6600
7.6258
3.3843
Image 4
4.2890
1.2558
5.1484
6.5282
1.2279
Image 5
4.4061
1.8448
5.5580
6.7848
2.7416
Image 6
4.7069
1.8610
5.8964
7.2198
3.1217
Image 7
4.9840
2.2702
5.9060
7.5486
3.4957
Image 8
4.9840
2.8858
7.5975
9.0504
4.9125
Image 9
5.6296
1.8776
7.3274
8.7363
4.2441
Image 10
4.4109
1.8439
5.1080
6.6907
2.3007
Image 11
4.2589
1.8322
5.0304
6.4672
2.1057
Image 12
4.6000
1.8588
5.4645
6.9777
2.5605
Image 13
4.8500
1.8636
5.5922
7.2857
2.9329
Image 14
4.7812
1.8452
5.2238
7.2293
2.4356
Image 15
5.1189
1.8585
5.8886
7.8711
3.0162
Image 16
4.4030
1.8225
4.8171
6.7702
1.7317
Image 17
4.6883
1.8329
5.0357
7.1325
2.0981
Image 18
4.9999
1.8478
5.4244
7.7206
2.3904
Image 19
4.5015
1.8503
4.6182
6.9186
1.5868
Image 20
4.6754
1.8570
4.6443
7.1054
1.7645
10.0000 9.0000
MSE Values
8.0000 7.0000 6.0000 5.0000
Gaussian Filter
4.0000
Laplacian Filter
3.0000
Wiener Filter
2.0000
Average Filter
1.0000
Median Filter
Im ag e Im 1 ag e Im 3 ag e Im 5 ag e Im 7 ag Im e 9 ag e Im 11 ag e Im 13 ag e Im 15 ag e Im 17 ag e 19
0.0000
Sample Images
Fig. 2 Comparison between the MSE values of different filters
66
S. Saxena et al.
Table 2 PSNR values calculated for 20 sample images Sample images Gaussian filter Laplacian filter Wiener filter Average filter Median filter Image 1
41.8811
45.3963
41.2613
40.1333
44.0885
Image 2
41.2891
45.3784
40.0671
39.4141
42.6634
Image 3
41.1932
45.4168
39.8960
39.3080
42.8362
Image 4
41.8072
45.4454
41.0141
39.9829
44.4336
Image 5
41.6903
45.4712
40.6816
39.8154
43.7508
Image 6
41.4034
45.4334
40.4250
39.5455
43.1868
Image 7
41.1550
45.4118
40.4179
39.3521
42.6954
Image 8
41.1550
45.3760
39.3241
38.5641
41.2177
Image 9
40.6261
45.3948
39.4813
38.7175
41.8529
Image 10
41.6855
45.4735
41.0483
39.8761
44.5122
Image 11
41.8378
45.5010
41.1148
40.0237
44.8969
Image 12
41.5032
45.4386
40.7553
39.6937
44.0476
Image 13
41.2734
45.4273
40.6550
39.5061
43.4578
Image 14
41.3355
45.4703
40.9510
39.5398
44.2648
Image 15
41.0391
45.4391
40.4307
39.1704
43.3362
Image 16
41.6933
45.5242
41.3030
39.8248
45.7460
Image 17
41.4207
45.4993
41.1102
39.5984
44.9125
Image 18
41.1412
45.4642
40.7873
39.2543
44.3461
Image 19
41.5973
45.4584
41.4861
39.7306
46.1255
Image 20
41.4326
45.4427
41.4616
39.6149
45.6647
48.0000
PSNR Values
46.0000 44.0000 42.0000
Gaussian Filter
40.0000
Laplacian Filter
38.0000
Wiener Filter
36.0000
Average Filter Image 19
Image 15
Image 17
Image 13
Image 9
Image 11
Image 5
Image 7
Image 3
Image 1
34.0000
Sample Images
Fig. 3 Comparison between the PSNR values of different filters
Median Filter
Assessment of Image Quality Metric …
67
Table 3 SSI values calculated for 20 sample images Sample images Gaussian filter Laplacian filter Wiener filter Average filter Median filter ‘Image 1
0.9211
3.6859
0.9184
0.9170
0.9026
Image 2
0.9210
3.3323
0.9173
0.9161
0.9026
Image 3
0.9212
3.1915
0.9174
0.9161
0.9038
Image 4
0.9213
3.4207
0.9180
0.9168
0.9046
Image 5
0.9232
3.5084
0.9194
0.9182
0.9052
Image 6
0.9239
3.5065
0.9210
0.9198
0.9060
Image 7
0.9246
3.4964
0.9217
0.9204
0.9065
Image 8
0.9246
3.1706
0.9221
0.9209
0.9076
Image 9
0.9247
3.0629
0.9242
0.9230
0.9060
Image 10
0.9264
3.3485
0.9263
0.9252
0.9034
Image 11
0.9275
3.4765
0.9274
0.9262
0.9040
Image 12
0.9281
3.4728
0.9280
0.9268
0.9039
Image 13
0.9296
3.4588
0.9295
0.9283
0.9052
Image 14
0.9274
3.4390
0.9276
0.9257
0.9038
Image 15
0.9265
3.2325
0.9266
0.9249
0.9035
Image 16
0.9248
3.4152
0.9250
0.9231
0.9024
0.9253
3.4235
0.9256
0.9235
0.9022
0.9246
3.2375
0.9248
0.9227
0.9024
Image 19
0.9441
3.6219
0.9343
0.9325
0.9020
Image 20
0.9439
3.6707
0.9340
0.9325
0.9008
0.9500 0.9400 0.9300 0.9200 0.9100 0.9000 0.8900 0.8800 0.8700
Gaussian filter Wiener Filter
Image 19
Image 15
Image 17
Image 13
Image 9
Image 11
Image 5
Image 7
Image 3
Average Filter Image 1
SSI Values
Image 17 Image 18
Sample Images Fig. 4 Comparison between the SSI values of different filters
Median filter
68
S. Saxena et al.
Table 4 SMPI values calculated for 20 sample images Sample images Gaussian filter Wiener filter Average filter Median filter Laplacian filter Image 1
1.0739
1.1091
1.1029
1.0500
12.5476
Image 2
1.0672
1.0970
1.0930
1.0400
13.0289
Image 3
1.0648
1.1049
1.0907
1.0540
12.9971
Image 4
1.0681
1.1009
1.0946
1.0570
12.4839
Image 5
1.0644
1.1026
1.0902
1.0510
12.6396
Image 6
1.0657
1.1012
1.0926
1.0580
13.0607
Image 7
1.0755
1.1206
1.1073
1.0610
13.4630
Image 8
1.0755
1.1196
1.1094
1.0640
13.8971
Image 9
1.0719
1.1062
1.1007
1.0620
13.5613
Image 10
1.0781
1.1119
1.1092
1.0670
12.5924
Image 11
1.0794
1.1129
1.1107
1.0690
12.5204
Image 12
1.0762
1.1091
1.1058
1.0680
12.9110
Image 13
1.0810
1.1250
1.1137
1.0710
13.2191
Image 14
1.0902
1.1245
1.1266
1.0810
12.9308
Image 15
1.0912
1.1234
1.1270
1.0820
13.0718
Image 16
1.0858
1.1209
1.1198
1.0750
12.3870
1.0909
1.1211
1.1264
1.0910
12.7549
1.0932
1.1255
1.1293
1.0810
12.9958
Image 19
1.1100
1.1533
1.1503
1.0990
14.1354
Image 20
1.1145
1.1551
1.1571
1.0970
14.5049
SMPI Values
Image 17 Image 18
1.1800 1.1600 1.1400 1.1200 1.1000 1.0800 1.0600 1.0400 1.0200 1.0000 0.9800
Gaussian Filter Wiener Filter Average Filter Median filter
Sample Images
Fig. 5 Comparative results of SMPI values for different filters
Assessment of Image Quality Metric …
69
value, low SMPI value and SSI value which is less than 1. Therefore, median filter is considered to be best over all the other filters.
2.5 Conclusion In this paper, various preprocessing filters such as median, average, Wiener, Laplacian and Gaussian filters are applied on 20 lung CT scan images. Later, comparison between PSNR, MSE, SMPI and SSI values is done in order to assess the performance of various filters. It is observed that median filter is more effective than other filters and showed promising results in eliminating the noise from the medical images.
References 1. Muthamil Selvi P, Dr. Ashadevi B (2020) Elimination of noise in CT images of lung cancer using image preprocessing filtering techniques. Intern J Adv Sci Technol 29(4s):1823–1832 2. Suren Makajua PWC Prasad, Abeer Alsadoona AK Singhb (2018) A. Elchouemic, Lung cancer detection using CT scan images. Sci Direct Proc Comput Sci 125:107–114 3. Sasikala S, Bharathi M, Sowmiya BR (2018) Lung cancer detection and classification using deep CNN. Intern J Innov Technol Expl Eng (IJITEE) 8(2S):259–262 4. Abdalla Mohamed Hambal, Dr. Zhijun Pei (2017) Faustini Libent Ishabailu: image noise reduction and filtering techniques. Intern J Sci Res (IJSR) 6(3):2033–2038 5. Abdillah B, Bustamam A, Sarwinda D (2016) Image processing based detection of lung cancer on CT scan images. Asian J 893:1–7 6. Parikh P, Ranade AA, Govind Babu K, Ghadyalpatil N, Singh R, Rangarajan Bharath, Bhattacharyya Gouri, Koyande S, Singhal M, Vora A, Verma A (2016) Hingmire, Sachin: lung cancer in India: current status and promising strategies. South Asian J Cancer 5(93):93–95 7. Panpaliya N, Tadas N, Bobade S, Aglawe R, Gudadhe A (2015) A survey on early detection and prediction of lung cancer. Int J Comput Sci Mob Comput 4(1):175–184 8. Zohair A Shamil, Ghazali S (2015) Latest methods of image enhancement and restoration for computed tomography: a concise review. Appl Med Inform 36(1):1–12 9. Sukhjinder Kaur (2015) Noise types and various removal techniques. Intern J Adv Res Electron Commun Eng (IJARECE) 4(2):226–230 10. Sivakumar S, Chandrasekar C (2014) A comparative study on image filters for noise reduction in lung CT scan images. Intern J Comp Sci Eng Inf Technol Res (IJCSEITR) 4(2):277–284 11. Azadeh Noori Hoshyar (2014) Adel Al-Jumaily, Afsaneh Noori Hoshyar: comparing the performance of various filters on skin cancer images. Proc Comp Sci 42:32–37 12. Vijaya G, Suhasini A (2014) An adaptive preprocessing of lung CT images with various filters for better enhancement. Acad J Cancer Res 7(3):179–184 13. Naitik P Kamdar, Dipesh G Kamdar, Dharmesh N Khandhar (2013) Performance evaluation of LSB based steganography for optimization of PSNR and MSE. J Inf Knowl Res Elect Commun Eng 02:505–509 14. Prasad DVR (2013) Lung cancer detection using image processing techniques. Intern J Latest Trends Eng Technol 3(1):372–378 15. Ansari A, Borse RY (2013) Image processing and analysis. Int J Eng Res Appl 3(4):1655–1658 16. Xin Wang, Linlin Ge, Xiaojing, Li. Evaluation of filters for Envisat ASAR speckle suppression in pasture area, ISPRS annals of the photogrammetry. Remote Sens Spatial Inform Sci 7:341– 346
70
S. Saxena et al.
17. Gupta G (2011) Algorithm for image processing using improved median filter and comparison of mean, median and improved median filter. Intern J Soft Comput Eng 1:304–311 18. Maini R, Aggarwal H (2010) A comprehensive review of image enhancement techniques. J Comput 2(3):8–13 19. Rafael C Gonzalez, Richard E Woods (2008) Digital image processing, Pearson prentice, 3rd edition, chapter 5, pp 311–360 20. Ramesh Jain, Rangachar Kasturi, Brian G Schunck (1995) Machine vision, image filtering, McGraw-Hill, Chapter 4, pp 112−139
A New Approach for Acute Lymphocytic Leukemia Identification Using Transfer Learning Saba Farheen Munshi and Chandrakant P. Navdeti
Abstract Leukemia a blood cancer, especially acute lymphocytic leukemia (ALL) is one of the most deadly diseases that affects adults and children, causing early deaths. Computerized software was then used to minimize chances of recommending improper medications. To prevent the deadly effects of leukemic disease, it become necessary to develop an automated and robust classification system. Existing methods related to leukemia classification were based on several segmentation techniques. In this paper, a fully automated classification system is proposed based on a pretrained transfer learning deep networks like AlexNet, and VGG16 architecture is applied. Out of which, AlexNet is used as classifier, whereas VGG16 is used as classifier and modified as feature extractor, where the feature vector extracted from VGG16 is given to support vector machine (SVM) classifier as input. Experimental comparison analyzes that our approach outperformed the performance against literature work. Experiments were performed on an ALL-IDB dataset which verified that the VGG performance is better than the AlexNet due to 100% accuracy of classification. Keywords Acute lymphocytic Leukemia · Transfer learning · Classification · Deep learning · AlexNet · VGG16 · ALL-IDB
1 Introduction A kind of cancer that can be induced by improper growth of the white blood cells is called as Leukemia. It arises in bone marrow. Figure 1a shows infected blood cells sample, whereas Fig. 1b shows normal blood sample. As per [1], leukemia has been categorized as chronic and acute. Acute leukemia spread rapidly that needs to be treated immediately, whereas chronic leukemia does not need to be dealt immediately. Depending on the type of cell involved, acute and chronic leukemia can be either myelogenous or lymphoblastic. Acute lymphocytic leukemia (ALL) is recognized as the primary focus for this research. A significant hematic disorder is lethal if left S. F. Munshi (B) · C. P. Navdeti Department of Information Technology, SGGS Institute of Engineering & Technology, Nanded, Maharashtra 431606, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_6
71
72
S. F. Munshi and C. P. Navdeti
(a)
(b)
Fig. 1 Sample blood smear image of ALL-IDB1 a infected and b non-infected [4]
untreated because of its fast dissemination through the bloodstream and other essential organs, and it primarily affect youngsters and individuals older than 50 years. Rapid identification of this disease is crucial for recovery in patients, particularly among children. Even in many diseases, symptoms of ALL are common, and therefore, the analysis is challenging. One phase of the diagnostic procedure involves analysis of the peripheral blood microscope. The work includes studying diseased white cells owing to the existence of a cancer. This project has been carried out for decades by professional technicians, who typically implement two key assessments: the classification and cell numbering. Curiously, morphological analysis comprises of an image only without a blood sample, which is optimal for cheap, accurate and remote testing. There are also few efforts for partial or fully automatic leukemia detection systems of image processing presented in literature [1–3]. In this paper, we worked on ALL-IDB, online data repository comprising of blood cell samples of healthy and leukemic persons. The remaining part of manuscript is arranged as follows: Sect. 2 covers relevant studies, Sect. 3 defines dataset with their two variants, ALL-IDB 1 and ALL-IDB 2, and method for collecting blood samples. Section 4 proposes methods for automated ALL Identification. Lastly, Sect. 5 contains the conclusions and future work.
2 Previous Work Some works have been suggested in particular to segmentation [5], to improve [6] or detect improper segmentation of white cells as proposed in [7]. In particular, the classification of cells by means of ANN and genomic operators is handled in [8]. The work showed in [9] reveals strategies for improving microscopic images by eliminating unnecessary background components, introduces a robust cell diameter estimation process and a modern methodology of self-adaptive segmentation for robust white cells identification. Outcomes revealed that feature extraction from white blood cells would have considerable classification accuracy of about 92%.
A New Approach for Acute Lymphocytic …
73
Several approaches have been suggested for detecting leukemia and implemented to classify two types of leukemia: AML and ALL. Ying Liu [10] introduced classification method into ALL and healthy blood cells implementation of ensemble learning over bagging dataset which has been used to reduce the influence of imbalanced dataset of total 76 images. The weighted F1-score of training and testing set were 0.84 and 0.88, respectively. Rahul Duggal [11] studied and recommends a machinebased learning method for segmenting affected WBC’s cell from microscopic image dataset. The technique proposed was composed of three stages: First is to remove nuclei from WBC blasts; this paper applied clustering of K-means in color space Lab channels ‘a’ and ‘b.’ Second is to identify clusters of nuclei that need splitting. In third step by applying a machine learning approach to classify pixels on the elevation of WBC blasts touch/overlap nuclei. Those elevation pixels get removed from the picture once detected seem to become segmented nuclei. Comparatively, proposed deep network belief approach for segmentation of these nuclei was observed to be best. Ahmad [12] has used several machine learning approaches for churn customers’ prediction in telecom field which predict customers those who are willing to leave the organization. Experiments were performed over decision tree (DT), random forest (RF), gradient boosted machine (GBM) and extreme gradient boosting (XGBoost). Whereas, highest accuracy result was obtained by applying XGBoost algorithm with an accuracy of 93.3%. Gupta [13] has proposed an implicit Lagrangian twin extreme learning machine (TELM) for classification via unconstrained convex minimization problem (ULTELMC) that remove the need of optimization toolbox used for solving quadratic programming problems (QPPs) further to examine the performance of proposed approach, several classification algorithms such as SVM, twin SVM (TWSVM), extreme learning machine (ELM), twin ELM (TELM) and Lagrangian ELM (LELM) were used and compared. Parashjyoti Borah [14] proposed constrained problem of optimization which was first revised into an unconstrained problem of minimization throughout this paper and afterward overcome by recursive convergent schemes to reduce the need of its corresponding dual formulation for solving quadratic programming problems (QPPs). Again to examine the performance of proposed approach, several classification algorithms such as SVM, twin SVM (TWSVM), extreme learning machine (ELM), twin ELM (TELM), random vector functional-link (RVFL) net, kernel ridge regression (KRR), constrained ELM (CoELM) and were used and compared their performances in terms of accuracy metric. Model output can degrade due to the issue of class imbalance learning (CIL). So, Hazarika [15] introduced a new model of SVM based on density weight for the problem of binary CIL (DSVM-CIL). In addition, to increase the training speed of DSVM-CIL, an improved 2-norm-based density-weighted least square SVM for binary CIL (IDLSSVM-CIL) was also introduced. Experimental results were then analyzed with SVM, least squares SVM, fuzzy SVM, improved fuzzy least squares SVM, affinity and class probability-based fuzzy SVM and entropy-based fuzzy least squares SVM for improvisation of generalized performance. Two effective entropy-dependent fuzzy SVM (EFSVM) variants have been proposed by Gupta D.
74
S. F. Munshi and C. P. Navdeti
R. [16]. One is an entropy-based fuzzy least squares SVM (EFLSSVM-CIL) and other was an entropy-based fuzzy least squares twin SVM (EFLSTWSVM-CIL). Fuzzy membership values are assigned based on sample entropy values for imbalanced class datasets for each sample. Experiments were performed on several imbalanced dataset and compared with various proposed algorithms like twin support vector machine (TWSVM), fuzzy TWSVM (FTWSVM), new FTWSVM for pattern classification (NFTWSVM), entropy-based fuzzy SVM (EFSVM), which show superiority of EFLSTWSVM-CIL. Another approach discussed by Amira Samy [17] implement a new type of leukemia detection based on the MFCC feature extraction and wavelet transform. Again discrete wavelet transformation (DWT) of blood sample signals were applied to remove extra MFFCs feature for identification process. So the steps are first convert 2D image to 1D signal then apply windowing over it with respect to fast Fourier transform (FFT) spectrum wrapped in Mel-filter sequence of banks according to Mel scale. Next step is to apply discrete cosine transform (DCT) followed by MFCC and then applying classification over MFCC vector. This paper followed five different classifiers like NBayes kernel, NBayes, LDA Quadratic, SVM Radial, and NCA for prediction of a normal cell and leukemic cell. NBayes kernel is considered to be better with the highest accuracy of 92.85%. Ahmed S. Negm [18] implemented a method where segmentation of the image was carried out automatically to monitor the occurrence of these repeated images (cells) also image enhancement used to optimize the accuracy. Using K-means clustering, blast identification in acute leukemia images were compared to LBG and KPE. The algorithm in K-means was preferable to the algorithms in LBG and KPE. Two separate approaches, neural network (NN) and decision tree (DT), had been used to perform classification. The model of the neural network yielded a better outcome with an accuracy of 99.74%, while the model of the decision tree was rapid with 88%. Prior studies concentrate on segmentation and the extraction of WBC functionality. Such features are hand-picked as inputs to standard supervised classification of models. Underreported in the identification of acute lymphoblastic leukemia is the reason for fine-grained automatic image recognition using CNN models. Richard Sipes [19] presented a procedure that includes hand-selecting attributes as inputs from cell images to a number of standard machine learning classifiers. Classifier used was KNN that results in an accuracy of 81%, whereas CNN is with 92%. Amjad Rehman [20] used image processing and deep learning approach for diagnosis of ALL which proposed a method of classifying normal images and ALL into its subtypes. Robust segmentation and deep learning techniques were used with the CNN to train the model on the bone marrow images to obtain accurate classification results. Experimental results have also been obtained and test results are compared with certain other classifiers—naive Bayesian, KNN and SVM. With the proposed approach, experimental results revealed 97.78% of accuracy. Sarmad Shafique [21] found an application of deep convolutional neural network (DCNN), in which pretrained AlexNet was used to detect and identify ALL into its L1, L2, L3 and normal subtypes. With data augmentation, they were able to achieve 99.50% accuracy for leukemia identification and 96.06% accuracy for the classification of its
A New Approach for Acute Lymphocytic …
75
subtypes. Therefore, the proposed approach was capable of achieving strong accuracy without the need for segmentation of microscopic images. Hence, literature work motivates us to overcome the drawback of existing work (an automated, fast, less complex, with high-performance system for detecting leukemia has to be developed).
3 Dataset and Background 3.1 The Dataset The images from the dataset were taken from a Canon Power Shot G5 camera with an optical research microscope. All dataset images are in JPG format, 24-bit color depth with a resolution of 2592 × 1944. The sample dataset images were captured from microscopes in scale from 300 to 500 magnifications. There are two different variants in ALL-IDB repository (ALL-IDB1 and ALL-IDB2) that can be publicly available [22].
3.1.1
All-Idb1
One can use ALL-IDB1 to evaluate the strengths of the models as well as preprocessing approaches for classification of image. That data includes 108 pictures taken in Sept. 2005. This includes nearly 39,000 blood samples, where oncologists labeled lymphocytes. Only those lymphoblast’s completely identified in the image are considered and classified [22].
3.1.2
All-Idb2
Here, set of images has developed to test the accuracy of classification system. The ALL-IDB2 is a set of regular and blast cells of concern cropped areas that belong to ALL-IDB1. It has 260 number of images, 50% of which are lymphoblasts. ALLIDB2 images have almost the same gray-level features as ALL-IDB1 images, without the size of the image [22].
76
S. F. Munshi and C. P. Navdeti
3.2 Classification Models 3.2.1
Data Augmentation
The dataset available for this study is limited. In order to maximize the dataset size and enhance the efficiency of our model, various data augmentation methods have been used, just like horizontal and vertical flips, rotation, and random brightness which we have performed by using ImageDataGenerator class, and results are shown in Fig. 2. So, data augmentation is a fundamental procedure in deep learning as in deep learning, we need a lot of information. In some cases, it is not possible to gather thousands or a large number of images, so data augmentation comes into the picture. It encourages us to increase the dataset size and add heterogeneity within the data.
Fig. 2 Images are augmented with zoom, rotation, height, width after applying ImageDataGenerator()
A New Approach for Acute Lymphocytic …
77
Fig. 3 Transfer learning architecture
3.2.2
Transfer Learning
Transfer learning is technique of learning where a model is learned and built for one task, and then reused for a second similar task as displayed in Fig. 3. This applies to the condition under which what was observed in one environment is applied to boost it in other environment. Transfer learning is typically introduced when a dataset is lesser than the actual one require to train the pretrained model. This paper introduces a new approach that utilizes model (Alexnet and VGG16) in which a standard dataset (ImageNet) [23] was first trained, and now has been reconfigured to learn or transfer functionality to be trained on a new dataset (ALL-IDB). So far as preliminary training is concerned, transfer learning helps one to continue with the learning features of the ImageNet dataset and fine tune these features and even the system models to suit the new dataset instead of initiating the learning cycle with random weight initialization on the data from scratch. Keras is used for facilitating pretrained model in CNN for transfer learning. To find an appropriate model, we examine the topology of CNN architecture, allowing for the classification of images via transfer learning. While examining and modifying the topology of the network as well as the feature of the dataset to assess which variables affect the accuracy of the classification, but with limited computational resources and time.
3.2.3
AlexNet
Krizhevsky et al. [24] developed AlexNet, which was used in imageNet [23] visual recognition. Figure 4 provides an example of the AlexNet architecture. So, it is splitted into five convolution layers and three fully connected layers, with eight
78
S. F. Munshi and C. P. Navdeti
Fig. 4 AlexNet architecture
compromise layers. The dimensions of input layer cannot exceed the predetermined width (W) and height (H) equivalent to 224 × 224, though three is that layer’s depth (D) as red, green and blue. The input image processed into the first layer of convolution. In the first layer, number of kernels (K) equivalent to 96 with a filter (F) size of 11 × 11 and a four pixels stride are used. Many of the convolutionary layers can include a number of , the output size of the padding (P). With the mathematical expression (W −F+2P) S+1 convolution layers can be determined. So, the size of the convolutional output will ≈ 55. The second convolutionary layer will have the input of 55 × 55 be (224−11+0) 4+1 × 96. With the use of two graphical processing units (GPUs), the work load would be 55 × 55 × 48 for each GPUs. In general, the dimensional reduction method has to be implemented with feature map. In this model, the pooling layer that follows the convolutionary layer is used to minimize the size of the feature map. Pooling can be average, max, sum, etc. Here × 55 × 96 ≈ the max pooling layer is preferred, though the size of the layer is 55 2 2 27 × 27 × 96. This layer of pooling is linked with 256 filter of size 5 × 5 × 48 and a 2-pixel stride. As discussed earlier, there are two GPUs, so the load for each GPU would be 27 × 27 × 128. The second layer output connected to 384 kernels makes up the third layer input. For two GPUs, the kernel size in each of this layer is 3 × 3 × 192. The fourth and fifth layers each have 384 and 256 kernels. In the fourth layer, each kernel size is of 3 × 3 × 192, whereas the kernel size for the fifth layer is 3 × 3 × 128. The third and fourth layers were formed without any layer of normalization or pooling, whereas the fifth layer has a layer of max pooling. Lastly, two layers are created which are fully connected. The input of these generated layers comes from the convolutionary third, fourth and fifth layers, whereas each layer is fully connected and has 4096 neurons.
A New Approach for Acute Lymphocytic …
3.2.4
79
Vgg16
VGG16 consists of 16 convolutionary layers, and its very uniform architecture tends to make it very appealing. Like AlexNet, it just has 3 × 3 convolutions with plenty of filters. The VGGNet weight configuration is available to the public and has been used as a baseline feature extractor in many other applications and challenges. Figure 5a shows the structure of VGG16 network and precisely summarized as follows: 1.
2.
3.
4.
5. 6.
The first two convolutional layers consist of 64 kernel filters and size of filter is 3 × 3. An input image is with 3 channels, i.e., RGB passed to Conv1 and Conv2 layers, so dimensions now change to 224 × 224 × 64. Then, the obtained output is forwarded to max pooling layer with stride of 2. The Conv3 and Conv4 layers are of 124 kernel filters and size of filter is 3 × 3. A max pooling layer with stride 2 follows these two layers, and the resulting output will be reduced to 56 × 56 × 128. The Conv5, Conv6 and Conv7 layers are with kernel size of 3 × 3. All three layers uses feature map of 256. A max pooling layer with stride 2 follows these layers. Conv8 to Conv13 are two sets of convolutional layers with kernel size of 3 × 3. All these sets of layers has 512 kernel filters. A max pooling layer with stride of 1 follows these layers. Conv14 and Conv15 layers are fully connected hidden layers of 4096 units followed by a softmax output layer (Conv16 layer) of 1000 units. Next, we used VGG16 model as classifier by replacing last dense layer of 1000 units to 1 unit as our prediction is binary.
3.2.5
Pretrained VGG16 CNN Model as Feature Extractor with Image Augmentation
We have implement VGG16 model, pretrained on ImageNet weights to extract features and feed the result to new classifier to classify images. We just need to include weights = ‘imagenet’ while calling VGG16 pretrained model also its important to set include_top = False to avoid loading the last fully connected layers. We need to introduce our own classifier as a pretrained model classifier, while our aim is to classify the image into two classes (normal or cancerous). After extracting lowlevel and high-level features like edges, lines, blobs and texture with feature vector size of 4096 as output from pretrained model will be used as input to new classifier to classify them into 2 classes as shown in Fig. 5b.
3.3 SVM Classifier We have chosen new classifier as SVM that tends to take input of 4096 feature vector generated by VGG16 pretrained model and separates it into 2 classes of dataset using
80
Fig. 5 a VGG16 architecture b VGG16 as feature extractor
S. F. Munshi and C. P. Navdeti
A New Approach for Acute Lymphocytic …
81
hyper plane having the common characteristics between the features of each instance. For better results, we have applied optimization parameters using GridSearch method. The optimal parameters used by GridSearch are C (regularization parameter), kernel (The model trains itself on the following kernels and gives us the best value between linear, poly, rbf, sigmoid and pre-computed values), gamma (Coefficient for the parameters of rbf, poly and sigmoid kernels). After applying the method, we got the optimal values as ‘C’: 1000, ‘gamma’: 0.1, ‘kernel’: ‘rbf.’
3.4 Random Forest Classifier We have chosen new classifier as RF that tends to take input of size (260,1000) feature vector generated by VGG16 pretrained model and classify it into two classes. For better results, we have applied feature selection step. For feature selection, we have used extra trees ensembling algorithm which works by generating a large number of extremely randomized decision trees, and predictions from the training dataset can be made by majority vote.
4 Proposed Methodology Following two classification approaches are created to differentiate microscopic pictures of healthy tissue and leukemia. Inspired by latest pretrained deep neural network approaches, transfer learning has been adopted for both methods. The first classification model comprises of only two stages: data augmentation and classification (see Fig. 6).VGG16 and AlexNet are employed in this work for classification of blood smear images.
Fig. 6 Block diagram of first classification model
82
S. F. Munshi and C. P. Navdeti
Fig. 7 Block diagram of first classification model
As in Fig. 7, the second model for classification composed of three stages: data augmentation, extraction of features and classification. As described earlier, data augmentation is used to resolve the shortage of data. In the extraction method, a pretrained VGG16 is used to retrieve a feature vector from individual image for further use in the classification to discriminate among normal and leukemic images. Some of well-known classifiers are used for the classification, such as SVMs and RFs. AlexNet and VGG16 have been used as proposed approach over Resnet50 and exception model, as these models work efficiently with less voluminous data. Our dataset contain less number of images to be trained, so we preferred AlexNet and VGG16 deep neural network which are not much complex compared to Resnet50 and exception which require large volumes of data for training, otherwise they tends to overfit.
5 Results Experiments were carried out using ALL-IDB datasets to validate an augmented image model. The performances of the proposed approach are calculated using five metrics: accuracy (1), precision (2), recall (3), F1-score (4), and kappa score. These metrics can be computed as: Accuracy =
(t p − tn) (t p + tn + f p + f n)
(1)
tp tp + f p
(2)
Pr ecision = Recall = F1 Scor e =
tp tp + f n
2(r ecall X pr ecision) (r ecall + pr ecision)
(3) (4)
A New Approach for Acute Lymphocytic …
83
Table 1 Accuracy analysis obtained for ALL-IDB database from proposed approach and certain works reported in the literature Work
Number of images
Accuracy (%)
Amjad Rehman [13]
108
97.78
Sarmad Shafique [17]
108
96.06
Proposed methodology
108
99.8
Table 2 Accuracy comparison obtained by the proposed approach over (a) ALL-IDB1 dataset (b) ALL-IDB2 dataset Work
Number of images
Accuracy (%)
AlexNet
108
50
VGG16
108
99.8
VGG16 + SVM
108
95
VGG16 + SVM + Grindsearch
108
99.8
VGG16
260
94.23
VGG16 + SVM + Grindsearch
260
78
VGG16 + Extratrees + RF
260
82.55
where tp, tn, fn and fp refer to true positive, true negative, false negative and false positive. All measures have high rate with the proposed method. The literature work and proposed work performances based on accuracy is shown in Table 1. We can claim the proposed approach worked much better than any other study listed in the literature while evaluating Table 1. Table 2 presents the result obtained by two versions of ALL-IDB datasets with varying size of feature vector. Table 2a represents VGG16 and SVM with parameters gives the best accuracy. Table 2b represents VGG16 and random forest with extra tree classifier gives the highest accuracy. Other classifiers also produce impressive results, but they are not sufficient to validate the approach followed. Figure 8a indicates accuracy rate with ALL-IDB1 and Fig. 8b indicates accuracy rate with ALL-IDB2 datasets achieved through the proposed architecture, i.e., VGG16 and having the highest result.
6 Conclusions and Future Scope In the proposed methodology, we have used several transfer learning CNN models to train and validate the augmented datasets such as AlexNet and VGG16. Where AlexNet is used as classifier. Whereas VGG16 model is used as classifier as well as employed to extract features and pass this feature vector as input to several other classifier such as SVM and RF. Experiments showed VGG16 classifier and VGG16 with SVM classifier along parameters estimated by GridSearch method was the
84
S. F. Munshi and C. P. Navdeti
Fig. 8 a Accuracy graph with ALL-IDB1 b Accuracy graph with ALL-IDB2
superior of all other classifiers. Apart from the work done toward this system, in future, we can extend leukemia identification to classify blood cell images among its subtype. Also, could implement such a software that be able to detect all kinds of cancer.
References 1. Goodell DJ (2000) Computer-assisted discrimination among malignant lymphomas and leukemia using immunophenotyping, intelligent image repositories, and telemicroscopy. IEEE Transactions on Information Technology in Biomedicine, pp 265–273 2. Kim KS (2000) Analyzing blood cell image to distinguish its abnormalities. Proceedings of the Eighth ACM International Conference on Multimedia (pp 395–397). Marina del Rey, California, USA: Association for Computing Machinery 3. Kovalev VA, Grigoriev AY-S (1996) Robust recognition of white blood cell images. Proceedings of 13th international conference on pattern recognition, pp 371–375 4. Aires LH (2018) Leukemia diagnosis in blood slides using transfer learning in CNNs and SVM for classification. Eng Appl Artif Intell 415−422 5. Nor Hazlyna H, Mashor MY, Mokhtar NR, Aimi Salihah AN, Hassan R, Raof RA (2010) Comparison of acute leukemia Image segmentation using HSI and RGB color space. 10th international conference on information science, signal processing and their applications (ISSPA 2010), pp 749–752 6. Badawy BP-S (2006) A high throughput screening algorithm for Leukemia cells. Canadian Conf Elect Comput Eng 2006:2094–2097 7. Lovell PB (2001) Method for accurate unsupervised cell nucleus segmentation. 2001 conference proceedings of the 23rd annual international conference of the IEEE engineering in medicine and biology society, pp 2704–2708 8. Scotti F (2005) Automatic morphological analysis for acute leukemia identification in peripheral blood microscope images. CIMSA. 2005 IEEE international conference on computational intelligence for measurement systems and applications, pp 96–101 9. Scotti F (2006) Robust segmentation and measurements techniques of white cells in blood microscope images. 2006 IEEE instrumentation and measurement technology conference proceedings, pp 43–48 10. Liu YA (2019) Acute lymphoblastic leukemia cells image analysis with deep bagging ensemble learning. bioRxiv
A New Approach for Acute Lymphocytic …
85
11. Duggal RA (2016) Overlapping cell nuclei segmentation in microscopic images using deep belief networks. Proceedings of the tenth Indian conference on computer vision, graphics and image processing 12. Ahmad AJ (2019) Customer churn prediction in telecom using machine learning in big data platform. J Big Data 2196–1115 13. Gupta PB (2020) Unconstrained convex minimization based implicit Lagrangian twin extreme learning machine for classification (ULTELMC). Appl Intell 1327–1344 14. Parashjyoti Borah DG (2019) Unconstrained convex minimization based implicit Lagrangian twin random vector Functional-link networks for binary classification (ULTRVFLC). Appl Soft Comput 15. Hazarika BG (2020) Density-weighted support vector machines for binary class imbalance learning. Neural Comput Appl 16. Gupta DR (2018) Entropy based fuzzy least squares twin support vector machine for class imbalance learning. Appl Intell 48:4212–4231 17. Samy AA (2017) A new approach for Leukemia identification based on Cepstral analysis and wavelet transform. Intern J Adv Comp Sci Appl 18. Kandil AS (2018) A decision support system for acute Leukaemia classification based on digital microscopic images. Alexandria Eng J 2319−2332 19. Sipes RA (2018) Using convolutional neural networks for automated fine grained image classification of acute Lymphoblastic Leukemia, pp 157–161 20. Rehman AA (2018) Classification of acute lymphoblastic leukemia using deep learning. Microsc Res Techn, pp 1310–1317 21. Tehsin SS (2018) Acute Lymphoblastic Leukemia detection and classification of its subtypes using pretrained deep convolutional neural networks. Technol Cancer Res Treatm 1533033818802789 22. Scotti RD (2011) All-IDB: the acute lymphoblastic leukemia image database for image processing. 2011 18th IEEE international conference on image processing, pp 2045–2048 23. Fei-Fei JR (2009) ImageNet: a large-scale hierarchical image database. 2009 IEEE conference on computer vision and pattern recognition, pp 248–255 24. Krizhevsky AA (2012) ImageNet classification with deep convolutional neural networks. Proceedings of the 25th International conference on neural information processing systems, vol 1, pp 1097–1105. Red Hook, NY, USA, Curran Associates Inc
A Hybrid Machine Learning Approach for Customer Segmentation Using RFM Analysis Poonam Chaudhary, Vaishali Kalra, and Srishti Sharma
Abstract Due to COVID-19 situation, online retailing (electronic retailing) for purchasing goods has recently increased which leads to the need of customer segmentation. Customer segmentation is done based on customers’ past purchase behavior and then divide them into different categories, i.e., loyal customer, potential customer, new customer, customer needs attention, customers require activation. This paper uses recency, frequency, monetary value (RFM) analysis and K-means clustering technique for grouping the customers. Further to enhance the efficiency of segmentation, a decision tree is used to create nested splitting (based on Gini index) inside the each cluster. The implementation of proposed hybrid approach is showing promising results for customer segmentation to take better management decisions. Keywords Online retail · Customer segmentation · RFM model · K-means clustering · Decision tree · Gini index
1 Introduction Online retailing (electronic retailing) means the efficient usage of the Internet to purchase goods and services by the customer. E-retailing is divided into 2 parts: • business-to-business • business-to-consumer In this paper, we are working upon business-to-consumer sales data to segment customers based on different categories, i.e., loyal customer, potential customer, new P. Chaudhary (B) · V. Kalra · S. Sharma The NorthCap University, Gurugram, Haryana, India e-mail: [email protected] V. Kalra e-mail: [email protected] S. Sharma e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_7
87
88
P. Chaudhary et al.
Table 1 Business-to-consumer and business-to-business E-retailing Business-to-consumer E-retailing
Business-to-business E-retailing
In this type of retailing, companies sell its products and goods to the customer directly with the help of Internet through their Website. This helps them to maintain good relationship with customers
In this type of retailing, companies sell its services to other companies. Wholesalers sell their goods and products directly from the manufacturing unit to the retailers
customer, customer needs attention; customers require activation using RFM analysis [1]. • Recency means how recent a transaction is. Recency is number of months since December when last transaction for the year is done by a customer. • Frequency is frequency of buying made by a customer in the year. • Monetary value is the total amount spent by a customer in the year covering all transactions for each customer. However, even the simplest of financial organization goes forward with at least one unsupervised machine learning to segment customers, the most commonly used method for grouping customers via K-means clustering. K-means is a very efficient and easy way of clustering similar things together [2, 3]. The only disadvantage is that parameter differences between points of different cluster should be really big; otherwise it becomes difficult for K-means to identify data points of different clusters as two different clusters. Going forward, we have used K-means as the method to cluster customers according to RFM which helps in identifying which customer should be in which category on these existing clusters of data points. Also, we have used elbow curve to compute the value of “K” by running K-means algorithms on different values of K and taking the minimum value before the saturation of cluster per category when plotted as a curve. The second method which we have used is supervised machine learning method, i.e., decision tree to segment the customers further. Decision tree splits the data with respect to a specific parameter. Decision tree is divided into two entities: decision nodes and leaves. The leaves are the final outcomes, and decision node is the entity from where the data split [4, 5]. This paper includes elbow and silhouette index for checking the strength and appropriate number of clusters (Tables 1 and 2).
2 Business Background and the Associated Data Khajvand et al. [6] have implemented RFM analysis for segmenting the customers in different groups and applied count item, resulted in new policies and strategies for retail companies. Christy et al. [7] have used K-means and C fuzzy means algorithms on transactional data segmented by RFM analysis. They concluded the research in terms of time, cluster compactness and iteration. Performance of customers from each
A Hybrid Machine Learning Approach for Customer …
89
Table 2 Customer category and definition in terms of RFM Customer category
General definition
Definition in terms of RFM
Loyal customer
This type of customer keeps coming back to purchase goods from the specific retail store
Recency should be minimum frequency and monetary value should be maximum
Potential customer
This type of customer has high potential to enter into loyal customer segment
Recency should be minimum; frequency and monetary value should be high but less than loyal customers’ frequency and monetary value
New customer
This type of customer is the fresh customer that just bought something from the retail store
Recency should be minimum frequency and monetary value should be minimum as he has just bought 1 or 2 products from the store
Customer needs attention
This type of customer has made some initial purchase but have not seen them since 5 to 6 months
Recency should be medium and having some frequency and monetary value which indicates that he has purchased some products earlier
Customer requires activation
This type of customer might have gone with the competitors for now
Recency should be high and having some frequency and monetary value which is less than customer’s needs attention’s frequency and monetary value
segment is missing in their work. Combination of RFM analysis with demographic information of customer has been used as feature set for further segmentation of customers in clusters by Sarvari et al. [8]. Do˘gan et al. [9] have done RFM analysis of 700,032 customers and clustered them compact clusters using K-means algorithm. They have not provided the strength of cluster and measure on number of clusters. The online-based retailing dataset has 8 factors as appeared in Tables 3 and 4, and it shows all the transactions occurred in years 2010 and 2011. The customer variable is primary for the business to provide important information that makes each costumer distinct, and then it does analysis to categorize customers and business to take better decisions.
3 Data Preprocessing To perform RFM model-based analysis, the original dataset needs to be preprocessed [11]. The fundamental advances associated in data preprocessing are as per following:
493414
C493415
C493426
493427
493427
493427
12
13
14
15
16
493414
8
11
493414
7
493414
493414
6
493414
493412
2
10
C493411
1
9
493410
0
Invoice
21682
21681
82483
22109
21527
21531
21527
35001G
37508
21533
21844
TEST001
21,539
TESTQQ1
StockCode
LARGE MEDINA STAMPED METAL BOWL
GIANT MEDINA STAMPED METAL BOWL
WOOD 2 DRAWER CABINET WHITE FINISH
FULL ENGLISH BREAKFAST PLATE
RETRO SPOT TRADITIONAL TEAPOT
RETRO SPOT SUGAR JAM BOWL
RETRO SPOT TRADITIONAL TEAPOT
HAND OPEN SHAPE GOLD
NEW ENGLAND CERAMIC CAKE SERVER
RETRO SPOT LARGE MILK JUG
RETRO SPOT MUG
This is a test product
RETRO SPOTS BUTTER DISH
This is a test product
Description
Table 3 Online-based retailing dataset, 2010 [10]
2010-01-04 10:41:00
−1
4
2
2010-01-04 10:43:00
2010-01-04 10:43:00
2010-01-04 10:43:00
2010-01-04 10:33:00
−3
4
2010-01-04 10:28:00
2010-01-04 10:28:00
2010-01-04 10:28:00
2010-01-04 10:28:00
2010-01-04 10:28:00
2010-01-04 10:28:00
24
12
2
2
12
36
2010-01-04 9:53:00
2010-01-04 9:43:00
5
−1
InvoiceDate 2010-01-04 9:24:00
Quantity 5
Price
4.95
9.95
5.95
3.39
7.95
2.1
6.95
4.25
2.55
4.25
2.55
4.5
4.25
4.5
CustomerlD
13287
13287
13287
16550
14590
14590
14590
14590
14590
14590
14590
123460
14590
12346
(continued)
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
Country
90 P. Chaudhary et al.
493427
493427
493427
17
18
19
Invoice
Table 3 (continued)
StockCode
21576
21577
18096C
LETS GO SHOPPING COTTON TOTE BAG
SAVE THE PLANET COTTON TOTE BAG
WHITE ROUND PORCELAIN TLIGHT HOLDER
Description
6
6
6
Quantity
2010-01-04 10:43:00
2010-01-04 10:43:00
2010-01-04 10:43:00
InvoiceDate
2.25
2.25
2.55
Price
13287
13287
13287
CustomerlD
United Kingdom
United Kingdom
United Kingdom
Country
A Hybrid Machine Learning Approach for Customer … 91
539993
539993
539993
539993
539993
539993
539993
539993
539993
539993
539993
539993
539993
539993
42481
42482
42483
42484
42485
42486
42487
42488
42489
42490
42491
42492
42493
42494
Invoice
85123A
22302
22303
22896
22898
22667
22961
20682
85099B
20718
22379
21498
21499
22386
StockCode
5
25
25
10
Quantity
6
10
WHITE HANGING HEART T-LIGHT HOLDER
COFFEE MUG PEARS DESIGN
COFFEE MUG APPLES DESIGN
PEG BAG APPLES DESIGN
CHILDRENS APRON APPLES DESIGN
RECIPE BOX RETROSPOT
12
6
6
6
8
6
JAM MAKING SET PRINTED 12
RED RETROSPOT CHILDRENS UMBRELLA
JUMBO BAG RED POLKADOT
RED RETROSPOT SHOPPER 10 BAG
RECYCLING BAG RETROSPOT
RED RETROSPOT WRAP
BLUE POLKADOT WRAP
JUMBO BAG PINK POLKADOT
Description
Table 4 Online-based retailing dataset, 2010 [10]
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
InvoiceDate
2.95
2.55
2.55
2.55
1.95
2.95
1.45
3.25
1.95
1.25
2.10
0.42
0.42
1.95
Price
13313
13313
13313
13313
13313
13313
13313
13313
13313
13313
13313
13313
13313
13313
CustomerlD
(continued)
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
United Kingdom
Country
92 P. Chaudhary et al.
539993
539993
539993
42495
42496
42497
Invoice
Table 4 (continued)
StockCode
22862
22458
22808
Quantity
LOVE HEART NAPKIN BOX
4
CAST IRON HOOK GARDEN 8 FORK
SET OF 6 T-LIGHTS EASTER 12 CHICKS
Description
2011-01-04 10:00:00
2011-01-04 10:00:00
2011-01-04 10:00:00
InvoiceDate
4.25
2.55
2.95
Price
13313
13313
13313
CustomerlD
United Kingdom
United Kingdom
United Kingdom
Country
A Hybrid Machine Learning Approach for Customer … 93
94
P. Chaudhary et al.
1.
Select some specific factors for data preprocessing from the given dataset. For our situation, four factors have been selected: Price, Quantity, Invoice Date, and CustomerID. Separate the dataset into 3 DataFrames according to years, i.e., 2010, 2011 with the help of InvoiceDate variable and then remove duplicate and null values from the DataFrames. Make new column month (purchase month) with the help of variable InvoiceDate to extract the month in which customer has made purchase of the product. Calculate recency by subtracting the variable month from 12 (December = 12) Calculate frequency by counting the number of purchase made per customer by variable CustomerID. Calculate monetory value by summing the value of variable price per CustomerID. Calculate minimum and maximum amount with the help of variable price by grouping CustomerID. Calculate first purchase month using new column that we have created, i.e., month by grouping CustomerID and then select value having maximum recency from that.
2.
3.
4. 5. 6. 7. 8.
After performing the given steps, a target dataset which contains the target features and variables has been generated .The target dataset and the features have been described in Tables 5 and 6, and statistics calculated have been described in Tables 7 and 8 for each year. Table 5 Data type description Variable name
Data type
Description
Invoice
Numeric
Invoice variable is a 6-digit number allotted to each transaction
StockCode
Numeric
Stock code variable is a 5-digit number combined with alphabets and allotted to each distinct product
Description
Numeric
Description variable is to describe or to name the product in words
Quantity
Numeric
Quantity variable is assigned to each product per transaction
InvoiceDate
DateTime
The date and time at each transaction were generated; for instance: 31/05/2011 15:59
Price
Numeric
Price variable is product price
CustomerID
Numeric
CustomerID variable is a 5-digit number assigned to each customer
Country
Numeric
Delivery address country; UK
2
0
0
2
1
1
2
7
1
1
1
2
1
3
9
9
12346
12347
12348
12349
12351
12352
12353
12355
12356
12357
12358
12359
12360
12361
12366
12368
18
3
19
89
85
41
165
83
22
20
18
21
102
37
71
40
79.05
5.07
78.90
278.22
380.42
160.29
1395.64
259.43
52.78
38.78
54.60
49.46
875.34
63.99
162.95
546.42
3
3
1
2
3
6
11
10
5
10
11
11
4
9
10
1
0 4
12363
10
2
0
0
1
1
7
8
7
1
10
1
3
0
11
12362
12361
12360
12359
12358
12357
12356
12355
12354
12353
12352
12350
12349
12348
12347
12346
23
274
10
129
254
19
131
19
13
58
4
95
17
73
14
151
2
53.17
1083.29
33.35
457.91
2225.11
157.21
438 67
188.87
54 65
261.22
24.30
2211.10
65.30
605.1
129.11
391.62
2.08
4
2
2
5
1
7
11
1
5
4
5
2
2
11
1
1
1
CustomerlD Recency frequency monetary_value firsf_purchase_month CustomerlD Recency frequency monetarv_value firsf_purchase_month
Table 6 RFM analysis of both dataset
A Hybrid Machine Learning Approach for Customer … 95
96
P. Chaudhary et al.
Table 7 RFM analysis 2010
Minimum
Maximum
Recency
0
11
Frequency
1
5762
Monetory value
0
39,807.85
First purchase month
1
12
2011
Minimum
Maximum
Recency
0
11
Frequency
1
7692
Monetory value
0
41376.33
First purchase month
1
12
Mean 2.75 96.205 373.68 5.17
Median 2 43 138.91 4
Table 8 RFM analysis Mean
Median
2.93
2
89.53
41
311.44 5.18
127.78 4
4 RFM Model-based Clustering Analysis With the prepared dataset in the wake of preprocessing, we need to distinguish whether consumers can be segmented meaningfully in terms of recency, frequency and monetary values. The K-means clustering algorithm is applied to find the optimal value of K for clustering the elbow curve method was implemented. When we apply K-means clustering algorithm on a dataset, it shows outliers or variables having incomparable magnitudes [12, 13]. The K-means clustering method required hit and miss trial for finding out the values of K that must be selected to find the best solution. Thus, so as to find the optimal solution of K, the elbow curve technique is used [14, 15].
Fig. 1 Number of cluster on 2010 dataset using elbow method
A Hybrid Machine Learning Approach for Customer …
97
Fig. 2 Number of cluster on 2011 dataset using elbow method
For the year 2010, the elbow curve, as shown in Fig. 1, determines the value of K as 2 or 3 whichever be more suitable. Similarly, for the year 2011, the elbow curve, as shown in Fig. 2, determines the value of K as 2. Silhouette coefficient for verifying the quality of cluster has been used and for cluster = 2 got the highest score of 0.71. For applying K-means clustering, we need to apply principle component analysis (PCA) so as to reduce the features that can be easily used in clustering process. The three important features such as recency, frequency and monetary value were taken, and reduced features were generated according to their correlation with each other. Accordingly, the clusters (Fig. 3a and b) were generated and their distribution has been shown for each year in Fig. 4a and b.
Fig. 3 a Clusters and their centroids on dataset 2010, and b clusters and their centroids on dataset 2011
98
P. Chaudhary et al.
Fig. 4 a Cluster distribution of dataset 2010, and b cluster distribution on dataset 2011
5 Enhancing K-Means Clustering Analysis Using Decision Tree From K-means clustering, we analyzed that cluster 2 is the most appropriate cluster and it has been selected as the optimal because it contains both customers: new as well as old and that too for both the years. To refine segmentation of the occurrence in the cluster, a decision tree [16, 17] is used to create nested splitting inside the cluster. These divisions form some sub-clusters inside cluster 2, and help to differentiate the consumers into some categories. For instance, for the year 2010, as shown in Fig. 5, the customers can be divided into categories such as frequency less than 91.5 and monetary value less than 333.9 and frequency less than 93 and monetary value less than 347 with initial splitting node as recency less than 2.5. Similarly, decision tree was constructed for year 2011 as shown in Fig. 6. Accordingly, the classes were assigned which divided the customers into the 5 sets of classes.
Fig. 5 Decision tree generation using Gini index on RFM features for dataset 2010
A Hybrid Machine Learning Approach for Customer …
99
Fig. 6 Decision tree generation using Gini index on RFM features for dataset 2011
6 Customer Centric Business Intelligence and Recommendations For each of the costumer groups, it is necessary to find out which of the items have been purchased by the customer of each group and how frequently they have purchased the product along with in which order they have purchased it. With the assistance of customer groups and the products they have purchased, the business can take better management decisions. Customer classification can also help in maintaining and engaging in business. Most of the customers were the organization who have been purchased products in higher quantity per transaction. Examining the frequency or how recent the products have been purchased would help the business for better understanding. Customer classification is done on the basis of different types of customers and their pattern to purchase the product which leads them to divide into some specific categories. Customer categories help to monitor the most to least potential customer and how much profit they earn from each category which helps to run the business more efficiently. Identifying appropriate categories for such predictions is extremely useful. It might enable the business to investigate different components that may influence client’s purchasing expectation and inclinations.
7 Concluding Remarks A contextual investigation that has been introduced in this paper is used to determine the methods in which customer-centric business for online retailers can be created by means of machine learning algorithms. The distinctive client bunches that have been described help the business for better comprehension as far as their benefit, and as needs be, take proper promoting techniques and choices for every client classification. It has been shown in this analysis that there are two most crucial
100
P. Chaudhary et al.
and time-consuming steps in the whole process includes information preprocessing and model understanding and assessment. Further research for the business includes categorize customers according to their buying patterns in terms of recency, frequency and monetary value into different categories that we have mentioned above; and to enhance retailer’s Website, so that they can track and capture customers’ shopping activities more accurate.
References 1. He X, Li C (2016) The research and application of customer segmentation on e-commerce websites. In: 2016 6th International conference on digital home (ICDH). IEEE 2016 2. Maulina NR, Surjandari I, Rus AMM (2019) Data mining approach for customer segmentation in B2B settings using centroid-based clustering. In: 2019 16th International conference on service systems and service management (ICSSSM). IEEE, 2019 3. Pakyürek M, Sezgin MS, Kestepe S, Bora B, Düza˘gaç R, Yıldız OT (2018) Customer clustering using RFM analysis. In: 26th signal processing and communications applications conference (SIU). Izmir, pp 1–4 4. Leiva RG et al (2019) A novel hyperparameter-free approach to decision tree construction that avoids overfitting by design. IEEE Access 7:99978–99987 5. Sarma VM, Abate AY (2014) A new decision tree approach to image data mining and segmentation. Int J Inf Technol Comput Sci Perspect 3(2):928. Market Customer Strategy Manage 14(2):130–142 6. Khajvand M, Zolfaghar K, Ashoori S, Alizadeh S (2011) Estimating customer lifetime value based on RFM analysis of customer purchase behavior: case study. Proc Comput Sci 3:57–63 7. Christy AJ, Umamakeswari A, Priyatharsini L, Neyaa A (2018) RFM ranking–an effective approach to customer segmentation. J King Saud Univ-Comput Inf Sci 8. Sarvari PA, Ustundag A, Takci H (2016) Performance evaluation of different customer segmentation approaches based on RFM and demographics analysis. Kybernetes 9. Do˘gan O, Ayçin E, Bulut ZA (2018) Customer segmentation by using RFM model and clustering methods: a case study in retail industry. Int J Contemp Econ Adm Sci 8(1):1–19 10. https://archive.ics.uci.edu/ml/datasets/online+retail 11. Sheshasaayee A, Logeshwari L (2018) Implementation of Rfm analysis using support vector machine model. In: 2018 2nd International conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC)I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (ISMAC), Palladam, India, pp 760–763 12. Maulani MR, Pane SF, Awangga RM, Wijayanti DA, Caesarendra W (2018) An analysis of customer agrotourism resort behaviour based on RFM and mean shift clustering. In: 2018 International conference on applied engineering (ICAE), Batam, pp 1–5 13. Safari F, Safari N, Montazer GA (2016) Customer lifetime value determination based on RFM model. Market Intell Plan 14. Bholowalia P, Kumar A (2014) EBK-means: a clustering technique based on elbow method and k-means in WSN. Int J Comput Appl 105(9) 15. Yuan C, Yang H (2019) Research on K-value selection method of K-means clustering algorithm. J Multidiscip Sci J 2(2):226–235 16. Chaudhary P, Agrawal R (2019) A comparative study of linear and non-linear classifiers in sensory motor imagery based brain computer interface. J Comput Theor Nanosci 16(12):5134– 5139 17. Chaudhary P, Agrawal R (2018) Emerging threats to security and privacy in brain computer interface. Int J Adv Stud Sci Res 3(12)
AMD-Net: Automatic Medical Diagnoses Using Retinal OCT Images Praveen Mittal
Abstract Eye is both a very fascinating as well as complicated organ of human body. The anatomy or retinal structure of eye if observed and analysed carefully may help in understanding and curing many diseases. The sole purpose of classification of diseases through the images is to generate different retinal layers and hence look them up for eye-related diseases. Since, the degree of complexity of retinal build of eye is very high. Moreover, the differences in the internal structure, image objects, are challenges that direct the results of stated methods towards negative directions. The methodology of the process begins with image enhancement, and removal of noise presents in the image which will improve the contrast of the image. But various methods have different accuracy in results, and neural network-based approach has a lot ground for fact-finding. Through this paper, we aim at classification of retinal disease with the help of deep learning. Keywords Retinal optical coherence tomography · Diabetic macular oedema · Drusen · Choroidal neovascularization · DenseNet · ResNet · ImageNet
1 Introduction Ophthalmology is evolving by many research scholars. Eyes being a major part of human body require advanced techniques for detection and treatment of various diseases - OCT being one such technique for imaging.
P. Mittal (B) G.L.A University, Mathura, Uttar Pradesh, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_8
101
102
P. Mittal
Handling of OCT images becomes a tedious task due to various issues such as noise in the image [1], less visibility and unnoticeable variation in intensity between different layers of Retina. OCT images are used to examine various eye diseases and to know the condition of the retinal layout. Therefore, it stings a very important role in the field of ophthalmology. Denoising is usually one of the major steps in classification preprocessing [2]. There are many convolutional neural network (CNN)-based algorithm as mentioned in [3] which was carried out on different imaging technique to differentiate medical images of pneumonia. We have hooked upon OCT images for this process as they have a huge relevance when it comes to generating a schematic that defines multiple tissues of eye, called the retina [4]. It is also used for examination, by imaging the eyes of the patients with various eye conditions such as diabetic macular oedema and diabetic retinopathy. Noise makes its way into the OCT images during acquisition [5]. Sensor and circuitry of a scanner or digital camera could also raise the problem of getting noise to the image. Further, film grain can also add to be a reason behind image noise. Never intended to be introduced in image, it degrades the quality of image. Hence, processing an image becomes necessary. Image processing is the procedure of improving the quality and information content of the original data. Image enhancement and restoration are amongst some relief approaches that are used to improve the quality of image. A latest research in [6] expresses application of deep learning in medical image processing.
2 Related Work There are number of classification work has been done for OCT images till date, but they are for classifying diseased eye from normal eye [7]. The following section describes the various works done in classification of retinal spectral domain OCT (SDOCT) images. M. Treder et al. proposed model in [3], which is based on machine learning for retinal SDOCT image categorization for various types of diseases that are related with retina with the help of dataset obtained from Heidelberg. Authors use Inception v3 model for deep convolutional neural network where starting layer got trained on ImageNet, and final layer got trained for taken dataset. Their work showed a good result for age-related macular degeneration diseases. Their work was only designed for age-related macular degeneration disease and normal eye images (Fig. 1). The next step that comes is segmentation. Segmentation is a process of dividing an image into regions [2]. This technique is mid-level processing technique. This further aims to segment the OCT images and work further on the results obtained. Here, this technique aims at diagnose various diseases that comprise the retina of the eye. This method needs to find the thickness of each layer after finding different
AMD-Net: Automatic Medical Diagnoses Using Retinal OCT Images
103
Fig. 1 Imaging of retina [8]
layers, in order to examine the eye for various diseases. Different layers may consist of different diseases which need to be diagnosed [4]. What adds to the problem is that the gradient amongst the layers is decent, and hence it further becomes a tedious task to segregate the layers. In this paper, AMD-net is used that works on cross validation which aims to figure the classification, hence generating the desired classification or clustering of images (Fig. 2).
3 Dataset Dataset taken from Kaggle is used in this proposed AMD-net model. In the taken dataset, there are four classes of images, named as diabetic pigment degeneration, retinal neovascularization, drusen and healthy eye. Total of 50,000 images over which convolutional neural network have been created.
104
P. Mittal
Fig. 2 Layers of retina [9]
4 Proposed Work This proposed convolutional neural network for retinal image classification is described below. Table 1 describes the kernel size of 3 × 3 with rectified linear unit as activation function for the strides rate of one. The two dimensional convolutional neural network with filter variation from size of eight to sixteen, thirty-two and sixty four is performed under the same size of batch matrices of 3 × 3. Two dimension Max pooling layer of size 2 × 2 performed the filtration of the output of one layer to input to the another layer in convolutional neural network. We add extra rows and columns to regain the size of image again. Conv2D(Mask = 8, subnet = (3, 3), progresses = 1, stuffing = ‘valid’, rectified linear unit as activation function, input_shape = size). Conv2D(Mask = 8, subnet = (3, 3), progresses = 1, stuffing = ‘valid’, rectified linear unit as activation function) MaxPooling2D(pool_size = (2, 2)). Conv2D(Mask = 16, subnet = (3, 3), progresses = 1, stuffing = ‘valid’, rectified linear unit as activation function) Conv2D(Mask = 16, subnet = (3, 3), progresses = 1, stuffing = ‘valid’, rectified linear unit as activation function,) MaxPooling2D(pool_size = (2, 2)). Conv2D(Mask = 32, subnet = (3, 3), progresses = 1, stuffing = ‘valid’, rectified linear unit as activation function) Conv2D(Mask = 32, subnet = (3, 3), progresses = 1, stuffing = ‘valid’, rectified linear unit as activation function). MaxPooling2D(pool_size = (2, 2)).
AMD-Net: Automatic Medical Diagnoses Using Retinal OCT Images
105
Table 1 Classification of retinal images with different class of diseases Cover (type)
Output shape
Param #
Two dimensional convolutional neural network@1 (0, 254, 254, 8)
80
Two dimensional convolutional neural network@2 (0, 252, 252, 8)
584
Two dimensional maximum pool@1
(MaxPooling2 (0, 126, 126, 8) 0
Two dimensional convolutional neural network@3 (0, 124, 124, 16)
1168
Two dimensional convolutional neural network@4 (None, 122, 122, 16)
2320
Two dimensional maximum pool@2
0
(MaxPooling2 (0, 61, 61, 16)
Two dimensional convolutional neural network@5 (0, 59, 59, 32)
4640
Two dimensional convolutional neural network@
(0, 57, 57, 32)
9248
Two dimensional maximum pool@3
(MaxPooling2 (0, 28, 28, 32)
0
Two dimensional convolutional neural network@7 (0, 26, 26, 64)
18,496
Two dimensional convolutional neural network@8 (0, 24, 24, 64)
36,928
Two dimensional maximum pool@4
(0, 12, 12, 64)
0
Cutting of values@1 values)
(None, 12, 12, 64)
0
Image curvature level@1 (flatten)
(None, 9216)
0
compressed@1 (compressed)
(None, 128)
1,179,776
compressed@2 (compressed)
(None, 4)
516
Total variables: 1,253,756 Adjustable variables: 1,253,756 Non-adjustable variable: 0
Convolutional neural network of two dimension (Mask = 64, kernel_size = (3, 3), strides = 1, padding = ‘valid’, activation = ‘ReLU’). MaxPooling2D(pool_size = (2,2)).
5 Experiment Result and Discussion See Fig. 3, Tables 2, 3, 4 and 5. Precision = Recall = F1 − Score = 2∗
TP TP + FP
(1)
TP TP + FN
(2)
precision*recall precision + recall
(3)
106
P. Mittal
Fig. 3 Results of this proposed method on retinal images (i) typical precision, (ii) prototypical harm Table 2 Values assigned to variables for proposed AMD-net Variable
Value
Cost
‘categorical cross-entropy’
Small matrix indexes size
250
Epoch
200
Premature bring to an end
10
Prototypical turnpike
null
Stuffing
‘True’
Optimization technique
‘diffGrad’
multiprocessing
Null
Table 3 True and false labelling of images OCT images True label
CNV
247
2
0
1
DME
5
241
1
3
DRUSEN
20
0
220
10
NORMAL
3
2
1
244
CNV
DME
DRUSEN
NORMAL
AMD-Net: Automatic Medical Diagnoses Using Retinal OCT Images
107
Table 4 Training time and validation time of proposed AMD-net Type
Time (s)
Training time
266000
Validation time
165
Per epoch training time is 3800 s
Table 5 Processing time and classification time of proposed AMD-net Type
Time (s)
Processing time
0.212
Classification time
1.161
Total time
1.373
Precision
Recall
F1-score
0
0.898
0.988
0.94
250
1
0.983
0.964
0.973
250
2
0.99
0.88
0.831
250
3
0.945
0.976
0.945
250
0.952
1000
0.922
1000
Accuracy Macro avg
0.954
0.952
Support
6 Conclusion The above results show the processing time of the proposed AMD-net method is 0.212 s which is less than the processing time of the previous research till now. Further, the classification time [10] of the proposed AMD-net method is 1.161 s which is 12 secs less than the time taken by ResNet [11] for on the same dataset [12]. So if we talk about the total time taken [13] to process the retinal OCT images for classification is 1.373 s, which is less than the time taken by ImageNet [14] and ResNet [15] on the same dataset.
References 1. Abelian J, Baker RM, Coolen FPA, Crossman RJ, Masegosa AR (2014) Classification with binary classifications from a nonparametric predictive inference perspective. Comput Stat Data Anal 71:789–802
108
P. Mittal
2. Chen M, Wang J, Oguz I et al (2017) Automated segmentation of the choroid in EDI-OCT images with retinal pathology using convolution neural networks. Fetal Infant Ophthalmic Med Image Analysis 10554:177–84 3. Pai SA, Hussain N, Hebri SP, Lootah AM, Dekhain MA (2014) Volcano like pattern in optical coherence tomography in chronic diabetic macular edema. Saudi J Ophthalmol 28:159 4. Worrall D, Wilson CM, Brostow GJ (2016) Automated retinopathy of prematurity case detection with convolutional neural networks 5. Hussain MA, Bhuiyan A, Ishikawa H, Smith RT, Schuman JS, Kotagiri R (2018)An automated method for choroidal thickness measurement from enhanced depth imaging optical coherence tomography images. Comput Med Imag Graph 63:41±51. https://doi.org/10.1016/j.compme dimag.2018.01.001, PMID: 29366655 6. Deepa N, Chokkalingam SP (2019) Deep convolutional neural networks (CNN) for medical image analysis. Int J Eng Adv Technol (IJEAT) 8(3S). ISSN: 2249-8958 7. Keel S, Lee PY, Scheetz J et al (2018) Feasibility and patient acceptability of a novel artificial intelligence-based screening model for diabetic retinopathy at endocrinology outpatient services: a pilot study. Sci Rep:4330 8. Brown JM, Campbell JP, Beers A (2018) Fully automated disease severity assessment and treatment monitoring in retinopathy of prematurity using deep learning. Proc Med Imaging Inf Healthcare Res Appl 9. Oliveira J, Pereira S, Gonçalves L, Ferreira M, Silva CA (2015) Sparse high order potentials for extending multi-surface segmentation of OCT images with drusen. In: 37th Annual international conference of the IEEE engineering in medicine and biology society (EMBC), pp 2952–2955. https://doi.org/10.1109/embc.2015.7319011 10. Sidibè D, Sankar S, LemaõÃtre G, Rastgoo M, Massich J, Cheung CY et al (2017) An anomaly detection approach for the identification of DME patients using spectral domain optical coherence tomography images. Comput Method Prog Biomed 139:109±117. https://doi.org/10. 1016/j.cmpb.2016.11.001 PMID: 28187882 11. Ngo L, Yih G, Ji S, Han JH (2016) A study on automated segmentation of retinal layers in optical coherence tomography images. In: 4th International winter conference on braincomputer interface (BCI), 1–2. https://doi.org/10.1109/iww-bci.2016.7457465 12. Yadav SS, Jadhav SM (2019) Deep convolutional neural network based medical image classification for disease diagnosis. https://doi.org/10.1186/s40537-019-0276-2 13. Mittal P, Bhatnagar C (2020) Automatic classification of retinal pathology in optical coherence tomography scan images using convolutional neural network. J Adv Res Dyn Control Syst 12(3):936–942 14. Mittal P, Bhatnagar C (2020) Detecting outer edges in retinal OCT images of diseased eyes using graph cut method with weighted edges. J Adv Res Dyn Control Syst 12(3):943–950 15. Mittal P (2020) Automatic segmentation of pathological retinal layer using eikonal equation. In: 11th International conference on advances in computing, control, and telecommunication technologies, ACT 2020, pp 43–49
Performance Interpretation of Supervised Artificial Neural Network Highlighting Role of Weight and Bias for Link Prediction Sandhya Pundhir, Varsha Kumari, and Udayan Ghose
Abstract Artificial intelligence includes various technologies: machine learning, neural network, deep learning, robotic, etc. Machine learning (ML) is a part of artificial intelligence in which machine is trained to learn without being explicitly programmed, and artificial neural network (ANN) is a popular model that is used for machine learning. The balanced combination of weight and bias plays a vital role in artificial neural network for error prediction. This paper is intended to provide a detailed study of different supervised artificial neural networks: feedforward ANN, multilayer perceptron, PatternNet, cascade feedforward network based on the different combination of weight and bias for link prediction. A comparative study has been done which highlights that the performance of ANN gets better as weight and bias initialization changes. Keywords Deep learning · Artificial neural network · Graph neural network · Link prediction
1 Introduction Today, artificial intelligence (AI) has become an integral part of the human life in terms of different application in various areas like engineering, education, banking, medical sector, business, nanotechnology, climate, agriculture, etc. Now, we can see the utilization of ANN in various fields like language translation, social Web data filtering, speech recognition, and many more. Machine learning, a broad area of artificial intelligence that is defined as the machine based on computer that has the ability S. Pundhir (B) · U. Ghose GGSIPU Delhi, Delhi, India U. Ghose e-mail: [email protected] V. Kumari GLA University, Mathura, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_9
109
110
S. Pundhir et al.
to learn and they have the capability to find patterns from the data in order to predict the specific output. Artificial neural networks are a part of artificial intelligence that involves different technologies like deep learning and machine learning. Artificial neural network (ANN) is an analyzing system inspired by the biological nervous system. Similar to our human brain, ANN has artificial neuron like the neurons in biological brains. These artificial neurons consist of a collection of nodes and each edge connecting these neurons can transmit a signal to other neurons. Characteristics of ANN that offers a better and generally used technology for universal function approximation in numerical models are self-learning, non-linearity, adaptability, fault tolerant, and efficient in input to output mapping [2]. Different ANN uses different methods for evaluation of a system, and these methods are accomplished with certain strengths and weaknesses [1]. As described above, the use of ANN is not limited to one field. It can be implemented in various fields depending on its architecture. Some of the popularly used ANN is described below. 1.
2.
3.
4.
Feedforward artificial neural network (FFANN): - In this system, the output from one layer feedforward way starting with one layer then onto the next layer. There are no feedback connections. It can have multiple layer with hidden layers. Feedforward neural networks are mostly used in the field of computer vision and facial recognition. Multilayer perceptron (MLP): MLP consists of three or more layers, and they are used to classify the data that cannot be separated linearly. In MLP, each node is fully connected to its following nodes in the next layer. MLP is effectively used in speech recognition and other machine translation technologies. PatternNet: - They are also a feedforward neural network. It is comprised of vectors of all zero value other than the class in which they are present. Default layers neurons size is 10. It can also classify inputs according to target classes. Cascade feedforward: - Cascade feedforward networks behave almost same as feedforward networks, except it includes a connection from the input layer and every previous layer to the successive layers. As two layered feedforward network has the ability to learn input–output relationship, similarly feedforward having more number of layers in hidden layer can also learn complex relationship more quickly.
The basic architecture of neural network is shown in Fig. 1. It consists of an arbitrary number of basic computing elements called nodes placed on parallel and independent layers. Each node is associated with some weight. The bias node is a special node added to each layer in the neural network, and it mostly stores the value as one. The bias node is analogous to the offset present in linear regression and is represented as y = mx + c where “m” is called slope and “c” is the coefficient of independent variable “x.” Neural networks can be used for different learning models. One of its uses is as classifier, the input nodes will take the input features and the output nodes will have output classes. And another use is as a function approximation based on input and output node where input node takes inputs and output node gives output.
Performance Interpretation of Supervised Artificial Neural Network …
111
Fig. 1 Basic architecture of neural network
Role of bias in the prediction can be very important. Bias permits line to move up and down to fit the prediction for the given data way better. The bias value is generally set to 1. So the role of bias is to help confirms the results produced are the best fits the incoming signal. A bias neuron is generally added to control the behavior of the layer. The major function of bias is to give a constant value to a node along with the other inputs received by the network node. The importance of bias value is that it enables the movement of activation function either to left or right, for training the neural network. In this paper, we try to find the link prediction using several ANN learning algorithms. Role of bias is highlighted here. We have taken the following supervised machine learning classifier for testing the accuracy of different models like MLP, FFANN, PatternNet, and cascade feedforward. This paper is organized as follows: Related work in related field is presented Sect. 2. The sources of data and the various evaluation approaches used in this paper have been discussed in the next Sect. 3. Section 4 depicts the results, and finally, Section 5 concludes the study and gives the recommendations.
112
S. Pundhir et al.
2 Related Work In [2], artificial neural network applications that give the taxonomy of ANN and application challenges, their contribution; compare the performances and critiques methods. Various applications of ANN techniques in different areas are computing, science, agriculture, mining, engineering, medicine, environmental, technology, climate, business, arts, and nanotechnology. Neural network is better in terms of accuracy, latency, performance, processing speed, fault tolerance, volume, and scalability as per [1–3, 19, 20]. In [3], author discussed that support vector machine leads in terms of accuracy with 76.3% that balances the tradeoff between model bias and model variance. In [4], proposed a supervised pertaining technique provides better accuracy in testing as compared with the he_normal initialization and the bias value is set to 0. In [5], the optimization of different feedforward ANN architecture-based hyper parameter selection is done. And a new cost function AvgNew has been proposed. In [6], along with the review of classification based on ANN, concluded that among all other neural networks, back propagation neural network algorithm proves better. In [7], presented an approach to back propagation training, emphasized on first-order adaptive learning algorithm and uses in the context of deterministic optimization based on bias value. In [8], authors proposed link prediction in social networks using supervised learning task and they conclude SVM performs better in performance measures. In [9], author developed a new approach to assess the product quality and acceptability by changing the combination of weights and biases. In [10], it provides a comparative study of data and logic-related distortion and bias is discussed by the authors. In [11], authors clarify the uses of term bias and show how to measure and visualize the statistical bias and variance of learning algorithms. Also discusses methods of reducing bias and variance. In [12], presented a comparative study of unsupervised and supervised learning models in higher education scenario. In [13], authors developed a predictive model using multilayer perceptron of artificial neural network with update of weight and biases. They concluded that genetic algorithm used predicted better in context of the water intake, split tensile, and slump properties. In [14], authors discussed a different approach of graph neural network that is effectively used to represent the learning process. In [15], authors presented a survey paper on GNN. In reference [16], author presents an analysis report on the uses and popularity of different types of supervised model from 1950 till now. According to the report, the most frequently used supervised models are linear regression, logistic regression, neural networks, decision trees, support vector machines. The analysis shows that highest popularity model is artificial neural network. However, the other machine learning models are used in different application fields. In reference [17], Anthony M. Zador presents the similarity and dissimilarity of supervised learning with the animal brains by learning the behavior of animals and categorized it as supervised
Performance Interpretation of Supervised Artificial Neural Network …
113
and unsupervised based on their genome and also a bias-variance tradeoff in machine learning algorithm has been shown. In reference [18], authors compared a new approach with existing approaches to produce empirical charts and equations used. Deep feedforward artificial neural network (DFFANN) is used and highlighted that tailoring of DFFANN outperforms simple ANN in [19]. MyAct an activation method is proposed in [20] highlighting statistical properties of data. In reference [21], authors present a review paper on the methods and applications of graph neural network. In [22], the average performance has been done over the different experiments based on bias and variance that can be used for generalization performance and also presented the several implications for the combination of bias and the variance.
3 Data Sources and Evaluation Methods In [2], artificial neural network applications that give the taxonomy of ANN and application challenges, their contribution; compare the performances and critiques methods. Various applications of ANN techniques in different areas are computing, science, agriculture, mining, engineering, medicine, environmental, technology, climate, business, arts, and nanotechnology. Neural network is better in terms of accuracy, latency, performance, processing speed, fault tolerance, volume, and scalability as per [1–3, 19, 20] (Fig. 2). The databases used are available on UCI repository. Datasets taken are the collection of vertices and their connection to interfacing vertices in case exists [19]. 3.1.
Datasets: The datasets used in this paper are as follows:
Fig. 2 Different weight and bias combinations on FFANN
114
S. Pundhir et al.
A: C.
B. C.
3.2.
elegans frontal: It is a collection of neurons and synapses in C. elegans that contains the number of nodes = 297 and number of vertices = 2148. karate: This is a collection of social relationship that contains 34 members of karate club which consists of 34 nodes and 78 edges. USAir: This is a collection of domestic flights of US, US airports considered as nodes and air travel route among them are considered as edges which consist of number of nodes = 332 and number of edges = 2126.
Evaluation techniques: - Several techniques that have been used for evaluation of above mentioned is discussed. The value returned in the above evaluation methods varies in the range 0 to 1. Value 0 indicates no match, and value 1 indicates the perfect match. Evaluation techniques used in this approach to measure the quality of results obtained by taking several similarity measures. They are negatively-oriented scores, which mean lower values are better.
The following error values have been taken for every matched record pair: 1.
Mean absolute error (MAE): It is calculated as difference between two continuous errors variables. Here, x and y are the observations that show the same phenomena. M AE = |t − y|/n
2.
Mean: In statistics, mean is the measure of average value of a set of data. Mean is calculate as given below where x is the set of data. Mean = 1/n x
3.
(4)
AvgNew: A modified cross-entropy method proposed in. Avg N ew = (−t ∗ log(y) + (1 − t) ∗ log(1 − y))/(ll ∗ log(ll))
6.
(3)
Elapsed time: It is the measure of the total time taken to generate the required output. Elapsed time should be less for better performance, and it is calculated as ET = FinishTime − StartTime
5.
(2)
Cross-entropy (CE): This method is used in optimization and used to calculate estimate small probabilities accurately as shown below: C E = −T ∗ log(Y ) − (1 − T ) ∗ log(1 − Y )
4.
(1)
AvgNn: A modified cross-entropy method proposed in [5].
(5)
Performance Interpretation of Supervised Artificial Neural Network …
115
Avg N n = (−t ∗ log(y) + (1 − t) ∗ log(1 − y))/log(ll)
(6)
4 Experimental Work and Results This proposed work has been implemented on MATLAB 2016 with the configurations: 1.3 GHz Intel Core i5, memory 4 GB, and 125 GB hard disk. The work performed for all the four types of neural network considered here. The following are the combination of weight and bias used with the help of MATLAB:I. II. III. IV.
The weights and biases of each layer i are initialized according initial initialization. All layers initialize with weight and bias default value = 0. Initial weight and bias values generated for a layer based on Nguyen-Widrow method and it will evenly distribute the neurons over the input space. Every weight (bias) in any layer is set to new values that are calculated according to rands initialization function that produces random values between 1 and 1.
Table 1 given below shows that IVth combination of weight and bias gives the best result for feedforward artificial neural network. Again from Table 2 and above graph Fig. 3, IVth combination of weight and bias gives the comparatively better result for multilayer perceptron network. In Table 3, it shows that although execution time for IVth combination is high, but it gives less error comparatively and Table 4 also highlights the IVth combination of bias and weight (Figs. 4, 5 and 6). Table 5 shows the analysis of IVth configuration on FFANN with parameters: MAE, CE, MSE, AvgNew, AvgNn, mean, and time. This can be observed from above table that IV combination of weight and bias has good results for all four ANN used here. Table 1 FFANN with various weight and bias combination results FFANN
II
III
IV
0.1114
0.129
0.095
0.095
CE
0.401
0.130
0.346
0.095
MSE
0.319
0.115
0.086
0.048
AvgNew
0.028
0.035
0.023
0.054
AvgNn
0.29
0.343
0.220
0.134
Mean
−0.109
0.126
0.093
0.09
Time
0.83
0.657
0.56
0.51
MAE
I
116
S. Pundhir et al.
Table 2 MLP with various weight and bias combination results MLP
I
MAE
II 0.5
III 0.5
0.503
IV 0.5
CE
0.5
0.5
0.502
0.5
MSE
0.25
0.25
0.392
0.25
AvgNew
0.071
0.071
0.1011
0.071
AvgNn
0.69
0.69
1.49
0.69
Mean
−0.497
−0.497
0.5
−0.497
Time
0.61
0.49
0.39
0.45
Fig. 3 Different weight and bias combinations on MLP
Table 3 PatternNet with various weight and bias combination results PatternNet
I
II
III
IV
MAE
0.158
0.133
0.187
0.003
CE
0.158
0.134
0.187
0.003
MSE
0.133
0.108
0.157
0.003
AvgNew
0.005
0.009
0.0142
Inf
AvgNn
0.05
0.087
0.140
Inf
Mean
0.05
0.013
0.024
0.003
Time
0.7
0.84
0.8
1.62
5 Conclusion Many variant of ANN is available with lot many parameter choices even for weight and bias initialization. This paper presented the work with four different types of
Performance Interpretation of Supervised Artificial Neural Network …
117
Table 4 Cascade FF with various weight and bias combination results Cascade FF MAE
I 0.056
II
III
0.118
0.122
IV 0.096
CE
0.1
0.118
0.122
0.096
MSE
0.261
0.095
0.098
0.048
AvgNew
0.024
0.029
0.031
0.014
AvgNn
0.235
0.245
0.303
0.134
Mean
-0.097
0.115
0.119
-0.093
Time
0.497
0.67
0.551
0.56
Fig. 4 Different weight and bias combinations on PatternNet
Fig. 5 Different weight and bias combinations on cascade FF ANN
118
S. Pundhir et al.
Fig. 6 Analysis of IVth configuration on FFANN with different parameters
Table 5 FFANN with IVth configuration on Karate and USAir dataset
Karate
USAir
MAE
0.123
0.144
CE
0.123
0.144
MSE
0.062
0.091
AvgNew
0.0227
Inf
AvgNn
0.171
Inf
Mean
−0.114
Time
1.45
−0.139 0.62
ANN-based four different combinations of weight and bias. Here, these combinations are used at different layers and at different stage of model implementation for analysis. A comparative analysis highlights that here random initialization of weight and bias at every layer and with initial initialization has shown better results compared with others combination of weight and bias.
References 1. Ayon D (2016) Machine learning algorithms: a review. Int J Comput Sci Inf Technol 7(3):1174– 1179 2. Abiodun OI, Jantan A, Omolara AE, Dada KV, Mohamed NA, Arshad H (2018) State-of-the-art in artificial neural network applications: a survey. Heliyon 4(11):e00938 3. Mohamed AE (2017) Comparative study of four supervised machine learning techniques for classification. Int J Appl 7(2) 4. Peng AY, Koh YS, Riddle P, Pfahringer B (2018) Using su- pervised pretraining to improve generalization of neural networks on binary classification problems. In: Joint European
Performance Interpretation of Supervised Artificial Neural Network …
5. 6. 7. 8. 9.
10. 11.
12. 13.
14. 15. 16. 17. 18. 19. 20. 21. 22.
119
conference on machine learning and knowledge discovery in databases. Springer, Cham, pp 410–425 Ghose U, Bisht U (2020) Performance evaluation of various ANN architectures using proposed cost function. In: IEEE conference ICRITO 4–5 June 2020 Saravanan K, Sasithra S (2014) Review on classification based on artificial neural networks. Int J Ambient Syst Appl (IJASA) 2(4) Magoulas GD, Vrahatis MV (2006) Adaptive algorithms for neural network supervised learning: a deterministic optimization approach. Int J Bifurcation Chaos 16(07):1929–1950 Al Hasan M, Chaoji V, Salem S, Zaki M (2006) Link prediction using supervised learning. In :SDM06 workshop on link analysis, counter-terrorism and security, vol. 30, pp 798–805 Gonzalez Viejo C, Torrico DD, Dunshea FR, Fuentes S (2019) Development of artificial neural network models to assess beer acceptability based on sensory properties using a robotic pourer: a comparative model approach to achieve an artificial intelligence system. Beverages 5(2):33 Fomin VV (2014) The shift from traditional computing systems to artificial intelligence and the implications for bias. Smart Technol Fundam Rights Dietterich TG, Kong EB (1995) Machine learning bias, statistical bias, and statis- tical variance of decision tree algorithms. Technical report, Department of Computer Science, Oregon State University Sathya R, Abraham A (2013) Comparison of supervised and unsupervised learning algorithms for pattern classification. Int J Adv Res Artif Intell 2(2):34–38 Awolusi TF, Oke OL, Akinkurolere OO, Sojobi AO, Aluko OG (2019) Performance comparison of neural network training algorithms in the modeling properties of steel fiber reinforced concrete. Heliyon 5(1):e01115 Xu K, Hu W, Leskovec J, Jegelka S (2018) How powerful are graph neural networks? arXiv: 1810.00826 Wu Z, Pan S, Chen F, Long G, Zhang C, Yu Philip S (2020) A comprehensive survey on graph neural networks. IEEE Trans Neural Netw Learn Syst Supervised Learning Models popularity (2020) Homepage: https://www.kdnuggests.com/ 2018/12/supervised-learning-popularity-from-past-present.html. Accessed 29 May 2020 Zador AM (2019)A critique of pure learning and what artificial neural networks can learn from animal brains. Nat Commun 10(1):1–7 Naderpour H, Kheyroddin A, Ghodrati Amiri G (2010) :Prediction of FRP-confined compressive strength of concrete using artificial neural networks. Compos Struct 92(12):2817–2829 Pundhir S, Ghose U, Bisht U (2019) Tailored feedforward artificial neural net- work based link prediction. Int J Inf Technol:1–9 Pundhir S, Ghose U, Bisht U (2020) Assessment of effectiveness of data dependent activation method: MyAct. J Intell Fuzzy Syst:665–677 Zhou J, Cui G, Zhang Z, Yang Y, Liu Z, Wang L, Li CC, Sun M (2018) Graph neural networks: a review of methods and applications. arXiv:1812.08434 Raschka S (2018) Model evaluation, model selection, and algorithm selection in machine learning. arXiv:1811.12808v2
Using Big Data Analytics on Social Media to Analyze Tourism Service Encounters Sunil Kumar, Arpan Kumar Kar, and P. Vigneswara Ilavarasan
Abstract Indian tourism is the fastest-growing section of tourism; thus, the research of it is reasonable. This study examines the Indian tourism trends by examining the tweets on the social media Website Twitter. Twitter is a massive platform for sharing ideas all over the world. A total of 7, 91, 804 tweets were identified after removing the extra and non-related tweets from the downloaded tweets. When tourists visit any place, they often share their experience regarding service encounters at the destination on the social media platform. In this study, we endeavor to find out the sentiment and theme of discussions over social media by using sentiment analysis and topic modeling techniques. For sentiment analysis, we divided India into five zones (north, west, east, south, and north-east). Our findings show that overall tourists are enjoying their visits to the destination. However, east and north-east zones of India are facing some negativity. The Indian Government, service providers, and stakeholders can use these findings to do future planning for Indian tourism growth. Keywords Big data analytics · Sentiment analysis · Topic modeling · Indian tourism · Social media
1 Introduction India is a big country where approximately 1.34 million people follow different religions, celebrations, and dialects. They are performing various tourism activities that India offers to domestic and additionally, to the world’s general population. We can classify this tourism into picturesque tourism, mountain tourism, business tourism, shoreline tourism to provincial tourism, therapeutic and medical tourism, history tourism, and religious tourism. Such diverse tourism classes attribute themselves to the climatic, geographic, and traditional variations found in India’s 28 states and 09 S. Kumar · A. K. Kar (B) · P. V. Ilavarasan Department of Management Studies, Indian Institute of Technology, New Delhi 110016, India P. V. Ilavarasan e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_10
121
122
S. Kumar et al.
union territories. In today’s scenario, the Internet has changed the way of life, views, and perception of life. In the early 90s, the active user on the Internet was 1% of the total population, and now, it has been increased to more than 4.13 billion in 2019 [1]. At present, this is covering around 52.95% of the world’s population. Cell phones and high-speed broadband connections are available at a lower cost in the market. Because of that, it is effortless to reach the Internet. Available information on the Internet is vast and limitless [2]. Online tourism indicates all such informational entities that are present on the Internet in terms of links and Webpages that relate to travel [3, 4]. Indian tourism plays an essential role in India’s economy. Indian tourism contributes nearly 9.2% of the country’s GDP and 8.1% of the employments [5]. In the travel and tourism competitiveness report 2019, India secured 34th rank out of 140 countries. The report classifies India’s tourism sector’s price competitiveness 13th out of 140 countries [6]. It indicates that India has good air infrastructure. Nowadays, social media is a popular platform for tourists to share their views and feedback about the tourism product [7–9]. Besides, social media provides a platform where the tourism industry and tourists can connect with each other regarding their feedback, recommendations, and services [10, 11]. Tourism data from social media like Twitter [12] help us to get insight into current trends in the tourism industry. India is using the Twitter platform to promote Indian tourism all over the world. In this study, sentiment analysis [13] and topic modeling [14] are proposed to get valuable and useful information from the Twitter data related to Indian tourism. This study aims to know how social media like Twitter helps domestic and foreign tourists who want to visit India’s tourist places, i.e., how the availability of information (destination management organization’s contents and user-generated contents) influences the tourist to see India’s tourist places? [15]. The remaining part of the study is as follows: Sect. 2 illustrates the previous works done in the area of tourism and social media. Section 3 demonstrates the research methodology of the research. Section 4 illustrates the findings and discussions of the results. Section 5 illustrates the conclusion of the study.
2 Literature Review Social media is such a main platform that has been adopted by the world’s population for communication in the twenty-first century [16]. In addition, most of the industries, including the tourism industry and their customers, have used this platform to show their views [17], opinions, and ideas [18]. Social media, like Twitter, is a free platform for everyone to share reviews and opinions. Therefore, Twitter is having an immense amount of data [19]. This factor influences us to do sentiment analysis on it. Generally, sentiment analysis is a technique to know the people’s perceptions, emotions, and ideas toward a product or service [20, 21]. There are three main categories (negative, positive, and neutral) used in sentiment analysis [22]. Sentiment
Using Big Data Analytics on Social Media …
123
analysis deals with the extraction of sentiment from the user-generated content [23]. User-generated content can be categorized as subjective (i.e., opinionated) or objective contents (i.e., factual) [24]. Online reviews and social media data tell about happiness, joy, sadness, delight, etc. [25]. For tourism improvements, subjective content plays a significant role. Many researchers have used sentiment analysis in the field of tourism to find trends in the destinations’ choice [26, 27]. Topic models [14] have many applications in the field of natural language processing. Much research has been published based on topic modeling techniques in the field of software engineering, social science, environmental science, etc. The researchers [28] performed a study on topic modeling techniques in the software engineering area to tell how topic modeling has been applied to repositories. They used 167 articles between December 1999 and December 2014 for the survey. The researchers [29] surveyed some applications on topic modeling and discussed some problems. The researchers [30] reviewed the practical approaches of text mining techniques in information systems and topic modeling with regression analysis. Remarkably, tourist satisfaction is playing an essential role in analyzing tourism services and products. Tourist satisfaction is a necessary part of the marketing of tourism products because it enhances the consumption of tourism products and services. It also increases the probability of choosing the same destination for the next visit and recommendations [31]. Researchers presented a theoretical model to show the relationship among tourist satisfaction, behavioral intentions, and perceived quality. Their results show that if satisfaction is positive, then the customer’s repurchase intentions will be high [32].
3 Research Methodology This research objective is to get the sentiment and theme of discussion of the tourists for Indian tourism. Twitter is the leading source for getting tourist’s discussions about Indian tourism. On Twitter, an immense amount of information is available to Indian tourism under the hashtag (#), @mention, and user profiles. For extracting these data, the study utilized the power of the R language.
3.1 Data Collection The data for the analysis are collected from April 2019 to January 2020 in India. Data were collected on state-wise by using the @mentions of state tourism profiles, hashtags of famous monuments, and the hashtag of famous cities for each state’s tourism in India. About 7, 91, 804 tweets were extracted from Twitter during the above mention period.
124
S. Kumar et al.
3.2 Data Preprocessing Data preprocessing is an essential process because data should be noise-free for the significant data mining process. Data preparation includes normalization, removal of stop-wards, punctuations, extra white spaces, URLs, and numbers from the extracted Twitter data. The normalization technique transforms the Twitter data into lowercase and generalizes the form of words. Most of the extracted tweets contain # (hashtags) and @ (mentions) tags that are not valuable for the mining process, so we removed it from the tweets. Retweets and duplicates tweets have also been removed from the tweets for useful sentiment analysis, because these tweets were not showing the new idea.
3.3 Data Analysis Methodology Sentiment analysis [33] is performed in collected data from Twitter to know the sentiment of the tourist toward Indian tourism destinations using a semantic approach and natural language processing (NLP) [34, 35]. The sentiment of the tourists is categorized into positive, negative, and neutral. For the sentiment categorization, “Syuzhet” package of R language is used. This package contains the NRC library, which categorizes the tweets into eight emotions (fear, disgust, anger, sadness, joy, surprise, trust, and anticipation) with positive and negative sentiments. Topic modeling [36] is performed on the data to know the theme of discussion among the tourists across India. Further, Latent Dirichlet Allocation (LDA) [37] has been used to find the 20 topics with 20 terms in the dataset [38]. Word cloud was constructed on the topics to find out the highest frequency of the term. Further, a network diagram was prepared to cluster the highly co-related terms in LDA output.
4 Results and Discussion To understand the tourists’ sentiments about India tourism, we applied social media analytics methods on collected tweets.
4.1 Insights from Sentiment Analysis First, we divided the complete dataset into five zones (north, east, west, south, and north-east) of India. We applied a sentiment analysis on a zone-wise dataset, which is basically tourist discussions across India about Indian tourism. Figure 1 shows the sentiments of the tourist’s conversations zone-wise toward the Indian tourism
Using Big Data Analytics on Social Media …
125
Zone wise Sentiment Analysis 60.00 40.00 20.00 0.00 Negative East
Neutral North
North-East
Positive South
West
Fig. 1 Positive, negative, and neutral sentiments across India
destinations in terms of negative, positive, and neutral sentiments. Figure 1 demonstrates that each zone is having a vast number of positive sentiments of the tourists toward Indian tourism as compare to negative and neutral sentiments. It indicates that tourists are enjoying their trips and satisfying by tourism services at the destination. Figure 1 also demonstrates that east and north-east zones of India face negativity about the destination compared to other zones of India. Figure 2 shows the zone-wise emotions with the tourist’s sentiments about tourism destinations. Figure 2 shows that trust, joy, and anticipation have high sentiment scores as compared to other emotions. That means tourism destinations from each zone are maintaining the trust of the tourists. Tourists are enjoying their trips and showing more involvement in organized events at tourism destinations. In Fig. 2, fear and sadness also have a little bit of high sentiment score. In east and north-east zones, tourists feel fear and are not satisfied by the tourism services at the destinations compared to other zones of India. This creates dissatisfaction in the tourist for tourism destinations due to the crime, unmanaged activities, and poor infrastructure. Figure 3 shows negative and positive words which have been used by the tourists in the discussions. Tourists are often using “beautiful,” “good,” “thanks,” “best,” “like,” and “love” words in the discussion over the social media. This is showing that tourists are feeling happy at the destination and enjoying the trip. Also, Fig. 3 is showing some negative words like “miss,” “death,” “attack,” “rape,” and “killed” Emotion Chart 30 25 20 15 10 5 0
East
North
North-East
Fig. 2 Sentiment category with emotions across India
South
West
126
S. Kumar et al.
Fig. 3 Positive and negative across India
which have been used by the tourists in the discussions. It shows negativity in the tourists toward tourism destinations.
4.2 Insights from Topic Modeling First, we applied LDA algorithms on extracted tweets for finding the latent topics. Figure 4 shows that 20 topics are enough to determine the central theme of discussion from the dataset. Therefore, we calculated 20 topics with 20 terms. Then, we prepared a word cloud of these 20 topics with 20 terms using R language. By observing Fig. 4 Word cloud of topics across all tweets used in India
Using Big Data Analytics on Social Media …
127
Fig. 5 Network diagram based on the output of the LDA algorithm
the word cloud, the tourists more often discussed hospitality, accommodation, and location of the destination. In the end, we prepared a network diagram on the output of the LDA algorithm. In LDA output, there are 20 topics with 20 terms. Therefore, the network diagram contains a cluster of these correlated terms. Figure 5 shows 12 clusters with some color scheme. Each color differentiates clusters from each other. Each cluster in the network diagram represents a theme in the discussions among the tourists.
5 Conclusion This study’s primary purpose is to determine the sentiments [39] and latent topics discussed among the tourists on social media platforms. These outcomes can be beneficial to the Indian tourism industry and the Indian Government as well. The government can enhance tourism services based on the sentiment analysis and topic modeling outputs for India’s tourism. Results indeed show that tourists are enjoying their visits to the destination, but at the other end, they are facing the problem of rape, crime, hospitality, and kidnapping. If the tourist is satisfied by the destination,
128
S. Kumar et al.
he will revisit the same place and recommend that others visit the same destination [32, 40, 41]. Besides, highly satisfied tourists will make fewer complaints about the destination. If the tourist feels that the service or quality is higher than the money, he pays rather than tourist satisfaction [42]. Although, this study finds the sentiment of the tourists about the destination. The study does sentiment analysis on text data written in the English language [43]. Nowadays, tourists are using some image icons to show their feeling toward tourism products. This field is unexplored right now and allows the researchers to do work on that. Acknowledgements This paper’s research is funded by the Department of Science and Technology (DST), sanctioned 7/01/2019. The authors, responsively, acknowledge the opportunity given by DST to conduct academic research on the tourism of India.
References 1. Statista (2020) Global number of internet users 2005–2019. https://www.statista.com/statistics/ 273018/number-of-internet-users-worldwide/. Accessed 04 Mar 2020 2. Kim DY, Lehto XY, Morrison AM (2007) Gender differences in online travel information search: implications for marketing communications on the internet. Tour Manag 28(2):423–433 3. Xiang Z, Wöber K, Fesenmaier DR (2008) Representation of the online tourism domain in search engines. J Travel Res 47(2):137–150 4. Kar AK (2015) Integrating websites with social media—an approach for group decision support. J Decis Syst 24(3):339–353 5. Annual research: key highlights 2019 (PDF). WTTC. Accessed 15 Mar 2019 6. Travel and tourism competitiveness report 2019 (PDF). weforum.org. Accessed 10 Sept 2019 7. Litvin S, Hoffman L (2012) Responses to consumer generated media in the hospitality marketplace: an empirical study. J Vacat Mark 8(2):135–145 8. Xiang Z, Gretzel U (2010) Role of social media in online travel information search. Tour Manag 31(2):179–188 9. Kar AK, Rakshit A (2015) Flexible pricing models for cloud computing based on group decision making under consensus. Glob J Flex Syst Manag 16(2):191–204 10. So KKF, King C, Sparks BA, Wang Y (2014) The role of customer engagement in building consumer loyalty to tourism brands. J Travel Res 55(1):64–78 11. Verhoef PC, Reinartz WJ, Krafft M (2010) Customer engagement as a new perspective in customer management. J Serv Res 13(3):247–252 12. Mohan R, Kar AK (2017) Demonetization and its impact on the Indian economy–insights from social media analytics. In: Conference on e-business, e-services and e-society. Springer, Cham, pp 363–374 13. Alaei AR, Becken S, Stantic B (2019) Sentiment analysis in tourism: capitalizing on big data. J Travel Res 58(2):175–191 14. Banks GC, Woznyj HM, Wesslen RS, Ross RL (2018) A review of best practice recommendations for text analysis in R (and a user-friendly app). J Bus Psychol 33(4):445–459 15. Blackshaw P (2005, June 28) The pocket guide to consumer-generated media. http://www.cli ckz.com/3515576 16. Lassen NB, Madsen R, Vatrapu R (2014) Predicting iphone sales from iphone tweets. In: Paper presented at the 2014 IEEE 18th international enterprise distributed object computing conference, Ulm, Germany
Using Big Data Analytics on Social Media …
129
17. Gretzel U, Yoo KH, Purifoy M (2007) Online travel review study: role and impact of online travel reviews: laboratory for intelligent systems in tourism, Texas A&M University. LIST 2(07):01 18. Arora D, Li KF, Neville SW (2015) Consumers’ sentiment analysis of popular phone brands and operating system preference using Twitter data: a feasibility study. In: Paper presented at the 2015 IEEE 29th international conference on advanced information networking and applications, Gwangju, Korea 19. Gupta S, Kar AK, Baabdullah A, Al-Khowaiter WA (2018) Big data with cognitive computing: a review for the future. Int J Inf Manag 42:78–89 20. Schmunk S, Höpken W, Fuchs M, Lexhagen M (2014) Sentiment analysis: extracting decisionrelevant knowledge from UGC. In: Xiang P, Tussyadiah I (eds) Information and communication technologies in tourism 2014. Springer, New York, pp 253–265 21. Liu B (2012) Sentiment analysis and opinion mining. Synth Lect Hum Lang Technol 5(1):1–167 22. Yang D-H, Yu G (2013) A method of feature selection and sentiment similarity for Chinese micro-blogs. J Inf Sci 39(4):429–441 23. Lakhiwal A, Kar AK (2016) Insights from Twitter analytics: modeling social media personality dimensions and impact of breakthrough events. In: Conference on e-business, e-services and e-society. Springer, Cham, pp 533–544 24. Feldman R (2013) Techniques and applications for sentiment analysis. Commun ACM 56(4):82–89 25. O’Leary D (2011) The use of social media in the supply chain: survey and extensions. Intell Syst Account Financ Manag 18(2–3):121–144 26. Prameswari P, Surjandari I, Laoh E (2017) Mining online reviews in Indonesia’s priority tourist destinations using sentiment analysis and text summarization approach. In: Paper presented at the 2017 IEEE 8th international conference on awareness science and technology (iCAST), Taichung, Taiwan 27. Philander K, Zhong Y (2016) Twitter sentiment analysis: capturing sentiment from integrated resort tweets. Int J Hosp Manag 55(2016):16–24 28. Chen TH, Thomas SW, Hassan AE (2016) A survey on the use of topic models when mining software repositories. Empir Softw Eng 21(5):1843–1919 29. Daud A, Li J, Zhou L, Muhammad F (2010) Knowledge discovery through directed probabilistic topic models: a survey. Front Comput Sci China 4(2):280–301 30. Debortoli S, Müller O, Junglas I, vom Brocke J (2016) Text mining for information systems researchers: an annotated topic modeling tutorial. Commun Assoc Inf Syst 39(1):7 31. Kozak M, Rimmington M (2000) Tourist satisfaction with Mallorca, Spain, as an off-season holiday destination. J Travel Res 38(3):260–269 32. Petrick JF, Backman SJ (2002) An examination of the determinants of golf travelers’ satisfaction. J Travel Res 40(3):252–258 33. Devitt A, Ahmad K (2013) Is there a language of sentiment? An analysis of lexical resources for sentiment analysis. Lang Resour Eval 47(2):475–511 34. Kar AK (2020) What affects usage satisfaction in mobile payments? Modelling user generated content to develop the “digital service usage satisfaction model.” Inf Syst Front 22:1–21 35. Chang WL, Chen YP (2019) Way too sentimental? A credible model for online reviews 36. Hannigan TR, Haans RFJ, Vakili K, Tchalian H, Glaser VL, Wang MS, Kaplan S, Devereaux Jennings P (2019) Topic modeling in management research: rendering new theory from textual data. Acad Manag Ann 13(2):586–632 37. Blei DM, Ng AY, Jordan MI (2003) Latent Dirichlet allocation. J Mach Learn Res 3:993–1022 38. Grover P, Kar AK, Davies G (2018) Technology enabled health—insights from twitter analytics with a socio-technical perspective. Int J Inf Manag 39. Tech Talk (2019) 5 great ways big data can help small businesses thrive. https://tech-talk.org/ 2019/10/29/5-great-ways-big-data-can-help-small-businesses-thrive/ 40. Gotlieb JB, Grewal D, Brown SW (1994) Consumer satisfaction and perceived quality: complementary or divergent constructs? J Appl Psychol 79(6):875
130
S. Kumar et al.
41. Filieri R, Alguezaui S, McLeay F (2015) Why do travelers trust TripAdvisor? Antecedents of trust towards consumer-generated media and its influence on recommendation adoption and word of mouth. Tour Manag 51:174–185 42. Business Fundas (2016) How luxury hotels can compete with rivals through a great digital marketing strategy. https://www.business-fundas.com/2016/how-luxury-hotels-can-competewith-rivals-through-a-great-digital-marketing-strategy/ 43. García-Pablos A, Duca AL, Cuadros M, Linaza MT, Marchetti A (2016) Correlating languages and sentiment analysis on the basis of text-based reviews. In: Information and communication technologies in tourism 2016. Springer, Cham, pp 565–577
Measuring the Efficiency of LPWAN in Disaster Logistics System Nitin Rastogi and Aravendra Kumar Sharma
Abstract In the context of natural disaster, the core of the whole system is the management of logistics for the disaster hit people. Communication is the main factor for the management of logistics. Here, integration is obtained using an emergency management communication mechanism in which one of a kind portal for extraordinary troubles is connected to every other through a network of statistics flows. There is always lack of network since when a disaster occurs all the cellular network towers and other basic communication medium goes down making it impossible to communicate with others. Our disaster management system will be responsible for the displaying on time status of the availability, shortage, and the movement of the materials. It is designed keeping in mind the critical situation at the time of disaster and managing operations just with a click of a button. The built system will really be helpful in coordinating with all the operations seamlessly in no time. This type of situation requires a real-time transfer of data consuming less power and should be handheld device friendly. We have proposed an advanced communication solution which can be achieved using the latest Semtech’s LoRa (short for long-range) is a spread spectrum modulation technique derived from chirp spread spectrum (CSS) technology. This technology is part of the low-power wide area network (LPWAN) which do not require towers as well as require very less base stations. This proposed LPWAN transceiver model is a device which operates on ISM band of frequency range 863–928 MHz (varying according to continents). The transreceiver section has a text and voice sending technique so that we can get the exact required information in lower bits with a range up to 10 km without access point or towers. A GPS module is used to find the location of the transport system and the location of the requester so that we can track them on a real-time basis. Keywords LPWAN · Disaster management · Logistics management · Disaster logistics · LoRa network · Web application N. Rastogi (B) · A. K. Sharma Department of Information Technology, Institute of Management Studies, Ghaziabad, India e-mail: [email protected] A. K. Sharma e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 G. Sanyal et al. (eds.), International Conference on Artificial Intelligence and Sustainable Engineering, Lecture Notes in Electrical Engineering 836, https://doi.org/10.1007/978-981-16-8542-2_11
131
132
N. Rastogi and A. K. Sharma
1 Introduction The need for disaster logistics frameworks has arisen as long as natural disasters, and their detrimental impact is known to the Indian subcontinent. The huge list of natural disasters includes floods; severe snowfall; wild/forest fires; storms/hurricanes; earthquakes; pandemics. As the computer processing has emerged with latest technologies such as Internet of things (IoT), cloud computing, the Internet of things (IoT), and communication has 5G cellular networks, it is now feasible to provide low-cost networks that cover a wide area network such as LPWAN such as SigFox, LoRa, Symphony link, weightless, “NB-IoT,” and “LTE-M2.” These technologies allow low-power sensors which can consume power as low as 25–30 mW, wide area client, and up to a 10 km test link for the operating areas for sensor networks. Organizations such as “Goonj” have been facing problem in communicating with the logistics system. They were also facing problem in getting latest requirements from the remote location where the problem of network as well as power remains down even for seven days continuous [1]. We are focusing on the communication aspects of the problem, in present communication method of wireless communication, communication towers play a major role. The key use of these towers is to promote mobile nodes and other wireless networking devices with signal stations and increase their reception. But during any natural disaster or calamity such as earthquakes and flooding, these towers are falling down, making it impossible to connect with others. During the movement of transportation, the driver goes out of the network and we are not aware about his whereabouts. So to provide a hindrance free communication for the transport as well as the representative, we go for a long-range transceiver module which acts as a stand-alone device. Our work is divided into two categories. (i) (ii)
Disaster logistics assistance system (DLAS) Communication for the logistics system.
2 Related Work This section describes some of the research works done related to measuring the efficiency of LPWAN in disaster logistics system. The work done on the different prospects of the problem is stated as follows.
2.1 Disaster Logistics Assistance System (DLAS) Theft In the present scenario, the disaster management works with the traditional system of human-based tracking system and sending information using ERP-based/mobilebased system. The new research work done in the logistics system is the incorporation
Measuring the Efficiency of LPWAN in Disaster Logistics System
133
of small text messaging system using ham radio/satellite communication. The system depends on the intervention of human being 24 × 7 so that the system can be tracked and data could be sent to the source as well as destination time to time. The organization faces problem in controlling the theft during the logistics system since the area is disaster effected so maintaining the check on the logistics is very difficult. The data need to be updated in real-time on the server so that we can track the vehicle as well as estimate how much time it will take to reach the destination. They prefer local people to help in watch the movement which is not possible 24 × 7 and also people cannot judge whether the theft has occurred in the logistics. Movement The organizations are having problem in controlling the movement of the supply since many vehicles move to same destination but with very less or half the capacity. Organization is not able to design a movement path so that supply can be collected from different locations for single destination, instead they send material from different location tone location and then send to the destination. Organizations are using old traditional way to collect all goods at one location from different centers and then transport it to the destination. In such type of process, even though, the good is near to the destination but still the good is first sent to source center from where the transport has to start. They also do not have proper information how much supply is available at different centers.
2.2 Communication for the Logistics System Presently, in Indian perspective, the organizations are using mobile communication, chatrooms and IM, in-person communication, and PSTN telephonic conversations category for communication medium. Since we are concentrating on the wireless communication, so we will talk about only the mobile communication. In Indian subcontinent, we are using multiple communication mediums for monitoring of logistics system. Some of the mediums are as follows. Amateur Radio Amateur radio, also known as ham radio, is the use of radio channel for non-profit exchange of messages and emergency notification purposes [2]. It is a government approved two-way radio service that plays a key role in the public safety during disasters. The best thing is that it offers connectivity even in the worst case condition with global connectivity. It has multiple disadvantages such as (a) Amateur radio operations is directly affected due to the weather and terrain conditions. (b) It requires skilled operators and it can only be operated on advanced schedule meetings. (c) Amateur radios need a better antenna and repeaters for the long distance.
134
N. Rastogi and A. K. Sharma
GSM Technology GSM technology is being used for communicating over large distance with the use of tower technology. It has large data capacity and provides a good coverage. In the event of a disaster, telecommunications companies face many difficulties in ensuring the continuous functioning of the network and in helping both the public and the government in their recovery and restoration efforts. In addition, telecommunications operators face infrastructural and technological problems in the area of network stability and restoration. Wi-Fi/WiMAX Wireless Fidelity (Wi-Fi) stands for IEEE 802.11x, short-range wireless networking technology. Wi-Fi is a technology that uses wireless digital technology to connect several personal computers to mobile devices (such as PDAs and smartphones) and other terminals. Wi-Fi works in the 2.4 and 5 GHz license-free spectrum of 802.11 protocol groups. WiMAX is working on world interactability for microwave access is a standard based on IEEE 802.16 microwave broadband wireless access metropolitan area technology (MAN) (2). WiMAX provides high-speed broadband access that can be used to connect 802.11x wireless hotspots to the Internet. WiMAX signal coverage is up to 50 km, this system can run data transmission across a range of 50 km at a very high. Terrestrial Trunked Radio (TETRA) These specifications define the general infrastructure for communication for mobile radio [3]. It replaces the old analog telephones and handheld radios used by the companies of the next century, such as manufacturing or oil and gas. TETRA was collaboratively developed over the new standard, these standards provides earlier technological features such as mobile radio, cell phones, pagers, and cellular data. TETRA is already used in many countries but it is still not approved in Indian subcontinent. If approved, TETRA has many benefits over all above technologies: All communications are digital and encrypted [4]. All modes of one-to-one, oneto-many, and many-to-many communications are available which is very crucial for emergency situations. It is possible to transfer data on the same network. Instant and fast one-to-many calls that are vital to emergency situations. The much lower frequency used (380 MHz versus 800/900 MHz) provides a longer spectrum, which in turn enables very high degrees of regional coverage. The rather limited reach of a few hundred feet makes Wi-Fi an alternative not to be considered for public and emergency services [5].
3 Proposed Scheme The most challenging component in the response to a catastrophic event is the local distribution of critical supplies at the points of distribution (POD) [6]. This activity is
Measuring the Efficiency of LPWAN in Disaster Logistics System
135
hugely complex in disasters, because of the large number of PODs to be established, manned, and supplied and the severe impacts on the transportation and distribution networks. We are working on this module since last three years with different parameters and conditions. We tried it for home automation, logistics, and weather station. The LoRa network was failure in the case of home automation since multiple electromagnetic field effect and not able to reach very interior parts of the home created a very big problem. Logistics and weather station are best implementation grounds for the LPWAN networks. The objective of the disaster logistics system is to develop a responsible information system that displaying on time status of the availability, shortage, and the movement of the materials. It is designed keeping in mind the critical situation at the time of disaster and managing operations just with a click of a button. Analyzing the real-life scenario problem related to disaster logistics system, we communicated with “Goonj” (NGO) which deals in disaster logistics module. The organization informed us that they face many problems related to theft, tracking of vehicles, receiving consignment at the receiver end. They also face problem in receiving requirements instantly from their representative at the disaster end. Analyzing all aspects in respect to human power as well as technological aspects, we designed a solution to solve all the problems using minimum resources and using single hardware only. We divided the problem into two major aspects, one is logistics and another is communication. We analyzed that instead of logistics the problem of theft and tracking of vehicle was the main concern. It was worst since the organization was dependent on the human information which was available from different locations but the authenticity of the information was under suspicious. We researched and using the latest IoT technology designed a system which could provide real-time information related to the goods uploading and downloading without any intervention of the human being, and the data were sent in real time to the cloud. This information was then shared to different stakeholders from source till destination. The system consists of latest UHF-based RFID system which can read tags from distance of 2 m, these tags are attached to the goods and can reused after the distribution of the goods. The reader reads the tags in time period of 30 s and matched it with the previous list which is stored in the Raspberry Pi-based local server. This information is later sent to the remote server using LoRa network. Difference in the two consecutive data will trigger an alarm about the theft, proving change in the data, and the GPS location of the goods. The GPS connected to the Raspberry Pi sends the real-time geographical location to the server also solving the vehicle tracking problem as discussed in Fig. 1. Once the goods are downloaded at the destination, it will be read by the handheld device with the Goonj representative who can then distribute to the disaster affected people, which will be automatically updated to the server. Communication is the toughest part to handle in the case of disaster and logistics. As per the researched by us in the last three years, we compared different technologies available in the Indian market, common technologies used for such type of conditions are GSM, Walkie Talkie, and Ham Radios. Well, all these technologies
136
N. Rastogi and A. K. Sharma
Fig. 1 Vehicle tracking system
have different limitations such as towers, one-way communication, and pre-booking for communication. In fact, the biggest problem was power consumption by all these technologies and also requirement of any human being to handle them. We researched and concluded to move to LPWAN communication which can be used for M2M and IoT-based devices. LPWAN consists of many latest wireless technologies SigFox, NB-IoT, LTE-M, LoRa, etc., but the most affordable and approved in India are NB-IoT and LoRa. We compared both the technologies with their hardware and concluded that to solve our purpose LoRa was the best suited technology. We implemented the LoRa network using the Arduino module as well as Raspberry Pi, we tested the tracking of vehicle as well as the distance, it could cover. On implementation using Dragino LoRa Kit, we were able to get signal up to 5 km (non-line of sight), this low distance was due to large industries around the testing area. We also implemented UHF RFID-based tag tracking which was successful and sent data to the server. According to the market survey and implementation in India by statistical department, the following growing market share has come up. Figure 2 describes the number of connections required in upcoming years.
3.1 LoRa Technology Unlike other revolutionary innovations that might be slow to gain global acceptance, Semtech’s LoRa technology is not a sign of future potential but is now available across the globe. With more than a hundred known cases of use (and growing) and more than one billion devices deployed on every continent in the world, LoRa
Measuring the Efficiency of LPWAN in Disaster Logistics System
137
Fig. 2 LPWAN network comparison
devices and the LoRaWAN® protocol are creating a smart planet. Industry analyst IHS market reports that 43% of all LPWAN links will be focused on LoRa by 2023. LoRa technology understands the promise of the Internet of things (IoT). LoRa Technology Base LoRa technology works on the chirp spread spectrum. An expansion spectrum approach uses wideband linear frequency modulated chirp pulses to encode data. A chirp is a sinusoidal sign whose frequency will increase or decreases over time (frequently with a polynomial expression for the connection among time and frequency. It is described in Fig. 3. In digital communications networks, chirp spread spectrum (CSS) is a spectrum spread technology that involves wideband vector frequency modulated chirp signals to encode data. A chirp is a sinusoidal signal whose frequency increases or decreases over time (often with a polynomial expression for the relationship between time and frequency) [7]. Fig. 3 Chirp spread spectrum (CSS)
138
N. Rastogi and A. K. Sharma
CSS is characterized by low-power consumption, fading resilience, and Doppler effects. CSS coupled with digital modulation techniques such as binary orthogonal keying (BOK), quadrature phase shift keying (QPSK), and differential phase shift keying (DQPSK) can provide better output at bit error rate (BER) [8]. Chirp spread spectrum (CSS) was initially suggested for communication by Winkler [8] and extended to digital communication by Berni [8]. CSS is now being implemented for sensor networks due to its low energy requirements, channel degradation tolerance by multi-path fading, and Doppler effects. IEEE 802.15.4 adopts CSS for low-rate wireless local area networks (LR-WPAN) for spectrum and freedom of movement. Implementation of LoRa Technology Figure 4 explains the process of monitoring the goods movement using UHF RFID technology. UHF RFID tags are attached to each package r item as required, these tags can be read by the RFID reader from a distance of 2 m. The data are sent to the RPi server which keeps track of all the data. On any changes, it sends the information with the GPS location that changes in the data. Figure 5 explains about the communication between the RPi, LoRa network, and information and finally sent off to the server using the MQTT protocol. This structure is at the destination location or at the LoRa gateway which is connected to the Internet. This module received the data from sender through the LoRa receiver and transmits it to the RPi, the data are either computed at RPi or it is sent to the cloud/server using the MQTT protocols. In Fig. 6, A represents data from GPS and UHF RFID received at Raspberry Pi, B represents data sent to AWS IoT cloud for analytics, C represents Amazon API gateway for communication with other modules local as well as cloud, D represents AWS database for IoT devices DynamoDB and Lambda, E represents display of data on the Website, and F represents display of data on the mobile devices. The data collected from the vehicle GPS, and UHF RFID modules are sent using LoRa tranreceiver to the LoRa gateway, which transmits it to the AWS IOT cloud. The processing of data from the vehicle, disaster representative, and the control center is being processed on the rules already designed. This data are then sent to the Website as well as personalized data to the mobile users according to their uses. Figure 7 represents the LoRa hat for Arduino (two numbers), one with GPS and other without. The other hardware is the LoRa hat for Raspberry Pi, the hat is connected to the Raspberry Pi and then it gets connected to the LoRa network. The LoRa network is generated and maintained by the LoRa gateway (white box in the image), this gateway is connected to the Internet. LoRa gateways are counterparties that enable sensing devices to transfer information to the cloud. Allowing coverage in crucial indoor/outdoor locations. The long-range wireless picocell gateway is an IoT service that enables the LoRaWAN® architecture for use in homes, businesses, and tower blocks.
Measuring the Efficiency of LPWAN in Disaster Logistics System
139
Fig. 4 Real-time monitoring of goods on the vehicle
4 Analysis and Simulation Communication is the key for disaster logistics system so we are focusing only on the technologies which provide better communication at low-power but high range. We have proposed low-power wide area network as the solution for logistics system and compared as well as tested one of the technology. LPWAN is low-power wide area network also known as LPWA network [9]. It is a modern form of radio communication used in various forms of wireless data communication. Machine to machine-based Internet of things solutions. Key features inherent in the era are the lengthy range of communiqué, low bit charge, and small energy finances of transmission.
140
N. Rastogi and A. K. Sharma
Fig. 5 LORA gateway
Other technologies which are in comparison with the LPWAN are as follows: ZigBee 802.15.4 and 3G/4G/5G. As per the disaster conditions, we need to have technology which can stand on the following factors (All values are out of 10 where 1 is low and 10 means high.) i. ii. iii. iv. v. vi. vii. viii.
Good geographical penetration (>6) Good range (>6) Less power consumption (