Quantization watermarking in the JPEG2000 coding pipeline [ ]

Citation preview

QUANTIZATION WATERMARKING IN THE JPEG2000 CODING PIPELINE

Peter Meerwald

Department of S ienti Computing, University of Salzburg, Jakob-Haringer-Str. 2, A-5020 Salzburg, Austria pmeerw osy.sbg.a .at

In this paper, we propose a blind watermarking method integrated in

Abstra t

the JPEG2000 oding pipeline. Prior to the entropy oding stage, the binary watermark is pla ed in the independent ode-blo ks using Quantization Index Modulation (QIM). The quantization strategy allows to embed data in the detail subbands of low resolution as well as in the approximation image. Watermark re overy is performed without referen e to the original image during image de ompression. The proposed embedding s heme is robust to ompression and other image pro essing atta ks. We demonstrate two appli ation s enarios: image authenti ation and opyright prote tion. Keywords:

Digital watermarking, quantization, QIM, dither modulation, JPEG2000, robustness, image authenti ation, data hiding.

1.

INTRODUCTION

The watermarking problem is to embed a message into multimedia data in a robust yet imper eptible way. Previous resear h [1℄ indi ates that signi ant portions of the host image, e.g. the low-frequen y omponents, have to be modied. This led to the development of transform domain watermarking. Re ently, many wavelet-based watermarking s hemes have been proposed for appli ations su h as opyright prote tion [2, 3℄ or image authenti ation [4, 5℄. Corvi proposed one of the rst wavelet-domain watermarking algorithms, simply by applying Cox's additive algorithm [1℄ to the approximation image of a wavelet de omposition. Another te hnique manipulating oe ients of the approximation image is Xie's [4℄ algorithm whi h quantizes the median oe ient of a 3  1 sliding window. Kim [6℄ utilizes DWT oe ients of all subbands in luding the approximation image to equally embed a random Gaussian distributed watermark sequen e in

the whole image.

Per eptually signi ant oe ients are sele ted by

level-adaptive thresholding to a hieve high robustness.

The energy of

the watermark is depending on the level of the de omposition to avoid per eptible distortion. Kundur [7℄ embeds a binary watermark by modifying the amplitude relationship of three transform-domain oe ients from distin t detail subbands of the same resolution level of the host image. Sele ted oe ient tuples are sorted and the middle oe ient is quantized to en ode either a zero or a one bit. An early attempt to integrate wavelet-based image oding and watermarking has been made by Wang [8℄ and Su [9℄.

While the rst

approa h was based on the Multi-Threshold Wavelet Code  (MTWC) [10℄, the later proposal builds on Embedded Blo k Coding with Optimized Trun ation (EBCOT) [11℄ whi h is also the basis for the up oming JPEG2000 image ompression standard. Both watermarking algorithms add a pseudo-random Gaussian noise sequen e to the signi ant oe ients of sele ted detail subbands. In this paper, we present a blind watermarking te hnique integrated in the JPEG2000 oding pipeline.

The watermark embedding and re-

overy pro ess is performed on-the-y during image ompression and de ompression. The omputational ost to derive the transform domain a se ond time for watermarking purposes an therefore be saved.

Our

design builds on the results of the previously proposed wavelet-domain watermarking algorithms mentioned above. However, in order to t the JPEG2000 oding pro ess, our watermarking system has to obey the independent pro essing of the ode-blo ks. Algorithms whi h depend on the inter-subband [7℄ or the hierar hi al multi-resolution [5℄ relationship

an not be used dire tly in JPEG2000 oding. Due to the limited number of oe ients in a JPEG2000 ode-blo k, orrelation-based methods [3, 8℄ fail to reliably dete t watermark information in a single independent blo k. Obviously, watermarking methods that require a

ess to the original image or referen e data for watermark extra tion are not suited as well  this pre ludes all the non-blind s hemes [2, 6℄. In se tions 2 and 3, we will review quantization watermarking and briey dis uss the JPEG2000 oding model. Our JPEG2000-integrated watermarking algorithm is proposed in se tion 4. Experimental results demonstrating two appli ation s enarios are presented in se tion 5.

2.

QUANTIZATION WATERMARKING Blind watermarking is the ommuni ation of information via multime-

dia data where the unmodied host data is not available to the watermark re eiver (see gure 1). The watermark

w

(en oding the message

m

) is

host image

distortion

x

message m

w

embedder

Figure 1

y

0

+

x

= + x

?

w

+

x

= + x

0

y

extractor

m

?

Blind watermark ommuni ation model.

added to the host image x: The resulting watermarked image denoted by x

0

has to be per eptual identi al (or at least similar) to the origi-

nal image. This limits the amount of modi ation the host image an undergo in the embedding pro ess. Ramkumar [12℄ points out that for blind watermark ommuni ation, the host data x has to be seen as noise ( alled self-noise), mu h like the distortion y that an result from image pro essing or watermark atta ks. However, ontrary to the distortion y; the host image x is known to the watermark embedder. Many previously proposed watermarking te hniques add a spreadspe trum signal to the host image.

Be ause the unmodied host im-

?

age an not be subtra ted from the re eived signal x , the performan e of these blind additive watermarking s hemes suers from host-signal interferen e when orrelating the watermark with the re eived data. However, watermark embedding based on Costa's result [13℄ has been proposed by Chen [14℄ and further analyzed by Eggers [15℄. The te hnique alled quantization index modulation (QIM) uses the watermark message as an index to sele t a parti ular quantizer from an set of possible quantizers.

The sele ted quantizer is applied to the host data to

en ode the watermark message. For the watermarking method proposed in this work, the watermark message is a sequen e of N binary values, mn

2 f0 1g ;

: We use Chen's

[14℄ dither modulation system with two pseudo-random dither ve tors dl of length L: The embedding rule to pla e one bit of watermark information into a host image ve tor x of length L is given by

( ; )=

s xi m

where Q


(
) denotes


(xi + di (m))

Q

( )

di m ;

s alar uniform quantization with step size


=0

: For

watermark extra tion, the minimum distan e between the re eived data x

? and its losest re onstru tion point, belonging to either

m

=1

; is al ulated,

?

m

= argm min k

x

?

( ?; )k

s x

m

:

m

or

source image

compressed image

forward wavelet transform

bitstream analyzer

quantization

entropy decoder

Figure 2

ROI scaling

entropy encoding

ROI de-scaling

dequantizer

codestream rate allocation

inverse wavelet transform

compressed image

source image

The JPEG2000 oding pipeline.

We have hosen the dither modulation system for embedding mainly be ause of its simpli ity  the performan e of our watermarking system might be further improved using more e ient QIM methods [14℄.

3.

JPEG2000 CODING The up oming JPEG2000 image oding standard [16℄ is based on a

s heme originally proposed by Taubman and known as EBCOT (Embedded Blo k Coding with Optimized Trun ation).

The major dier-

en e between previously proposed wavelet-based image ompression algorithms su h as EZW [17℄ or SPIHT [18℄ is that EBCOT as well as JPEG2000 operate on independent, non-overlapping blo ks whi h are

oded in several bit layers to reate an embedded, s alable bitstream. Instead of zerotrees, the JPEG2000 s heme depends on a per-blo k quadtree stru ture sin e the stri tly independent blo k oding strategy pre ludes stru tures a ross subbands or even ode-blo ks. These independent ode-blo ks are passed down the  oding pipeline shown in gure 2 and generate separate bitstreams.

Transmitting ea h bit layer or-

responds to a ertain distortion level.

The partitioning of the avail-

able bit budget between the ode blo ks and layers (trun ation points) is determined using a sophisti ated optimization strategy for optimal rate/distortion performan e. The main design goals behind EBCOT and JPEG2000 are versatility and exibility whi h are a hieved to a large extent by the independent pro essing and oding of image blo ks [19℄. The default for JPEG2000 is to perform a ve-level wavelet de omposition with 7/9-biorthogonal lters and then segment the transformed image into non-overlapping

ode-blo ks of no more than

4096

oe ients whi h are passed down

the oding pipeline.

4.

WATERMARK EMBEDDING The watermark embedding stage is invoked after quantization and

region-of-interest (ROI) s aling and prior to entropy oding (see gure

3). At that point, ea h ode-blo k transports signed integer oe ients that have been normalized: the most signi ant bit (MSB) arries the sign bit and the remaining bits represent the absolute magnitude of the

oe ient. We have to distinguish between ode-blo ks belonging to either the approximation image (LL subband) or the detail subbands (LHj ; HLj ; HHj subbands, where j = 1 : : : J is the de omposition level). The nest resolution subbands an not be used to en ode information reliably. In the rst ase, i.e. ode-blo ks belonging to the approximation image, we apply an embedding te hnique similar to Xie's [4℄ approa h. We slide a non-overlapping w  1 running window over the entire ode-blo k. At ea h window position, one bit of watermark information is en oded using the quantization embedding te hnique des ribed in se tion 2. The size of the embedding window determines the oding rate. Given a grays ale image of size 512  512; the watermark information that an be
512 1 bits. Typi al values embedded in the approximation image is 512 22 J
w of w range from 2 to 8: For the ode-blo ks being part of one of the detail subbands, we have to use a larger embedding window be ause the energy is mu h lower. At least 256 oe ients are quantized to en ode one watermark bit. Thus, if the size of the ode-blo k allows, we an split it into several sub-blo ks to in rease the embedding apa ity. The large magnitude oe ients represent edge and texture information. The human visual system (HVS) is less sensitive to hanges in these regions, therefore we want to exploit this hara teristi to maximize the watermark strength. To keep the im plementation simple, a non-linear s aling fun tion f (x) = sign(x)
x ; > 1 is applied to all ode-blo k oe ients. The s aling parameter is hosen in a level-adaptive way between 6:5 and 5 . We obtain a more uniform oe ient representation sin e the high peaks in the oe ient distribution are redu ed. This way, we an use simple uniform s alar quantization (as before) and still put more watermark energy in the image regions the HVS is less sensitive to. After quantization, the inverse s aling fun tion f 1 is applied to derive the watermarked ode-blo k.


5.

RESULTS

We ondu ted our experiments with the JJ2000 implementation1 of the JPEG2000 veri ation model (VM). The modularized ar hite ture of the JJ2000 software allowed to easily integrate our watermarking module. If not noted otherwise, we use the default oding parameters for our experiments. To demonstrate the robustness and apa ity of our watermarking method, a watermark with 85; 194 and 383 bits was em-

code-block

ROI scaling

quantization

approximation image embedding

watermark embedding

detail subband embedding

window selection

sub-block selection

quantization embedding

non-linear scaling quantization embedding inverse scaling watermarked code-block

entropy encoding

Figure 3

The watermark embedding pro ess in the oding pipeline.

image

window size

sub-blo k size

apa ity

PSNR

Lena

4 2 2

64 32 16

85 194 383

32:05 31:45 32:09

Fishing Boat Goldhill

Table 1

Embedding parameters and the orresponding bit apa ity for three

512512

gray-s ale images, together with the resulting PSNR.

bedded in the

512  512

gray-s ale images Lena, Fishing Boat and

Goldhill, respe tively. Figure 4 shows the watermarked images Lena and Goldhill together with their dieren e images. The ee t of our simple s aling fun tion is learly visible: the edges ontain more watermark energy than smooth regions.

The dierent watermark apa ities

were a hieved by hoosing the embedding parameters from table 1. The resulting PSNR is also given. For opyright prote tion, we embed a binary message that identies the owner of the image. The dither ve tors are kept se ret to prote t the watermark.

The normalized orrelation result of the re overed versus

the embedded message is depi ted in gure 5. The watermarked images were subje ted to JPEG and JPEG2000 ompression with varying om-

pression parameters (top row).

To simulate image pro essing atta ks,

2 onvert

the images were blurred and sharpened using the ImageMagi k program (bottom row).

The results indi ate our watermark survives

the atta ks, but additional error- orre tive oding is required to a hieve perfe t re overy of the embedded information. The tamper dete tion appli ation requires a fragile watermark that breaks in order to indi ate the areas that have been manipulated.

At

the same time, however, the watermark should be robust against unintentional distortion, e.g.

aused by lossy image ompression.

Figure

6 (a) shows the Fishing Boat image, watermarked with a sequen e of

3 and

all-zero bits. Next, we manipulated three regions using the GIMP

JPEG ompressed the image with default quality; see gure 6 (b) and the dieren e image, highlighting the hanges ( ). The dete tion results of our watermarking s hemes are depi ted in gure 6 (d). The mali ious tampering has been dete ted and lo alized while the distortion due to JPEG ompression did not raise a false alarm. One oe ient in the approximation image of the wavelet domain orresponds to a blo k of pixels in the spatial domain. In order to a hieve good spatial resolution for our tamper dete tion example, we had to limit the wavelet transform to three de omposition steps. Therefore, we an authenti ate pixel blo ks of size

88

individually. Sin e the watermark

onsists of sequen e of zero bits, we ould use sliding window dete tion [20℄ in horizontal, diagonal and verti al dire tion in the approximation image. The tamper dete tion results from the three dire tions were a

umulated and ontribute to the brightness of the tamper dete tion image of gure 6 (d).

6.

CONCLUSION We demonstrated our watermarking s heme an be integrated in the

JPEG2000 oding pro ess and dis ussed some of the limitations. A novel embedding algorithm based on QIM and suitable for watermarking independent JPEG2000 ode-blo ks was proposed whi h allows blind watermark re overy during image de ompression. We investigated a opyright prote tion and an image authenti ation appli ation and provided robustness as well as apa ity results. Future work will try to improve the performan e of the embedding method and onsider ROI oding.

Notes

http://jj2000.epfl. h.

1.

The JJ2000 sour e is available for download at

2.

The ImageMagi k programs are at

3.

The GNU Image Manipulation Program is available at

http://www.simplesystems.org/ImageMagi k. http://www.gimp.org.

Figure 4

The watermarked images Lena and Goldhill, together with their dier-

en e images.

0.8

0.8

0.6

0.6 Correlation

1

Correlation

1

0.4

0.4

0.2

0.2 Image Lena (85 bits) Fishing Boat (194 bits) Goldhill (383 bits)

Image Lena (85 bits) Fishing Boat (194 bits) Goldhill (383 bits)

0

0 90

80

70

60

50

40

30

20

10

2

1.8

1.6

1.4

JPEG compression quality

1.2

1

0.8

0.6

0.4

0.2

JPEG2000 compression rate (bpp)

(a) JPEG

(b) JPEG2000

0.8

0.8

0.6

0.6 Correlation

1

Correlation

1

0.4

0.4

0.2

0.2 Image Lena (85 bits) Fishing Boat (194 bits) Goldhill (383 bits)

Image Lena (85 bits) Fishing Boat (194 bits) Goldhill (383 bits)

0

0 10

20

30

40

50

60

70

80

blur factor (percent)

( ) blur

Figure 5

90

10

20

30

40

50

60

70

80

90

sharpening factor (percent)

(d) sharpen

The robustness against JPEG ompression (a) and JPEG2000 ompression

(b), blurring ( ) and sharpening (d). Measured using normalized orrelation between the re overed and embedded message.

Figure 6

(a) watermarked image

(b) manipulated image

( ) dieren e image

(d) tamper dete tion

The watermarked Fishing Boat image (a) and the tampered version (b).

The manipulations are highlighted after default JPEG ompression in the dieren e image ( ). The manipulated regions dete ted by the algorithm (d).

Referen es [1℄ I. J. Cox, J. Kilian, T. Leighton, and T. G. Shamoon, Se ure spread spe trum watermarking for multimedia, in Pro eedings of the IEEE International Conferen e on Image Pro essing, ICIP '97, vol. 6,

pp. 1673  1687, (Santa Barbara, California, USA), O tober 1997. [2℄ M. Corvi and G. Ni

hiotti, Wavelet-based image watermarking for

opyright prote tion, in S andinavian Conferen e on Image Analysis, SCIA '97, (Lappeenranta, Finland), June 1997.

[3℄ M. Barni, F. Bartolini, V. Cappellini, A. Lippi, and A. Piva, A DWT-based te hnique for spatio-frequen y masking of digital signatures, in Pro eedings of the 11th SPIE Annual Symposium, Ele troni Imaging '99, Se urity and Watermarking of Multimedia Contents, P. W. Wong, ed., vol. 3657, (San Jose, CA, USA), January

1999. [4℄ L. Xie and G. R. Ar e, Joint wavelet ompression and authenti ation watermarking, in Pro eedings of the IEEE International Conferen e on Image Pro essing, ICIP '98, (Chi ago, IL, USA), 1998.

[5℄ H. Inoue, A. Miyazaki, and T. Katsura, Wavelet-based watermarking for tamper proong of still images, in Pro eedings of the IEEE International Conferen e on Image Pro essing, ICIP '00, (Van ou-

ver, Canada), September 2000. [6℄ J. R. Kim and Y. S. Moon, A robust wavelet-based digital watermark using level-adaptive thresholding, in Pro eedings of the 6th IEEE International Conferen e on Image Pro essing, ICIP '99,

p. 202, (Kobe, Japan), O tober 1999. [7℄ D. Kundur, Improved digital watermarking through diversity and atta k hara terization, in Pro eedings of the ACM Workshop on Multimedia Se urity '99, pp. 53  58, (Orlando, FL, USA), O tober

1999. [8℄ H.-J. Wang and C.-C. J. Kuo, An integrated approa h to embedded image oding and watermarking, in Pro eedings of IEEE ICASSP '98, (Seattle, WA, USA), May 1998.

[9℄ P.-C. Su, H.-J. Wang, and C.-C. J. Kuo, Digital watermarking on EBCOT ompressed images, in Pro eedings of SPIE's 44th Annual Meeting: Appli ations of Digital Image Pro essing XXII, vol. 3808,

(Denver, CO, USA), July 1999. [10℄ H.-J. Wang and C.-C. J. Kuo, High delity image ompression with multithreshold wavelet oding (MTWC), in SPIE's Annual meeting - Appli ation of Digital Image Pro essing XX, (San Diego, CA,

USA), August 1997.

[11℄ D. Taubman, High performan e s alable image ompression with EBCOT,

IEEE Transa tions on Image Pro essing 9, pp.

1158 

1170, July 2000. [12℄ M. Ramkumar and A. N. Akansu, Self-noise suppression s hemes for blind image steganography, in

dia Systems and Appli ations II,

Pro eedings of SPIE: Multime-

vol. 3845, (Boston, MA, USA),

September 1999.

IEEE Transa tions on Information Theory 29, pp. 439  441, May 1983.

[13℄ M. H. M. Costa, Writing on dirty paper,

[14℄ B. Chen and G. W. Wornell, Prepro essed and postpro essed quantization index modulation methods for digital watermarking, in

[15℄

Pro eedings of IS&T/SPIE's 12th Annual Symposium, Ele troni Imaging 2000: Se urity and Watermarking of Multimedia Content II, P. W. Wong, ed., vol. 3971, (San Jose, CA, USA), January 2000. J. J. Eggers and B. Girod, Quantization watermarking, in Pro eedings of IS&T/SPIE's 12th Annual Symposium, Ele troni Imaging 2000: Se urity and Watermarking of Multimedia Content II, P. W. Wong, ed., vol. 3971, (San Jose, CA, USA), January 2000.

[16℄  JPEG2000 part 1 nal ommittee draft version 1.0, te h. rep., ISO/IEC FCD15444-1, Mar h 2000. [17℄ J. M. Shapiro, Embedded image oding using zerotrees of wavelet

oe ients,

IEEE Transa tions on Signal Pro essing 41, pp. 3445

 3462, De ember 1993. [18℄ A. Said and W. A. Pearlman, A new, fast, and e ient image ode

based on set partitioning in hierar hi al trees, in IEEE Transa tions on Cir uits and Systems for Video Te hnology, vol. 6, pp. 243

 250, June 1996. [19℄ M. Charrier, D. S. Cruz, and M. Larsson,  JPEG2000, the next

millennium ompression standard for still images, in Pro eedings of the IEEE International Conferen e on Multimedia & Computing Systems, ICMCS '99, vol. 1, pp. 131  132, (Floren e, Italy), June 1999.

[20℄ J. J. Eggers and B. Girod, Blind watermarking applied to image

Pro eedings of the IEEE International Conferen e on A ousti s, Spee h, and Signal Pro essing, ICASSP '01, (Salt authenti ation, in

Lake City, UT, USA), May 2001.