IEEE AFRICON 2009 23 - 25 September 2009, Nairobi, Kenya

A Digital Image Watermarking Scheme Based onLocalised Wavelet Coefficients' Dependency

Ruth Dili BusePhysics DepartmentKenyatta University

PO BOX 43844Nairobi, Kenya

rdbuse@ku.ac.ke

Abstract- A wavelet based robust image watermarkingscheme in which a binary image is embedded in thehorizontal and vertical quadrants of a wavelet transformedimage is proposed. The algorithm exploits the inter­coefficients relationship within a DWT block of apredetermined size. Only one wavelet coefficient isperturbed slightly in order to embed a single bit of thewatermark and also to produce a watermarked image ofhigh perceptual quality. The embedded DWT coefficientsblock is selected by the use of a secret key in order toenhance the security of the scheme. The retrieval process isblind and only requires the knowledge of the key. Results,obtained by MATLAB simulation have been used toconfirm the feasibility and effectiveness of the proposedalgorithm against common attacks such as lossy imagecompression, low pass filtering, additive white gaussiannoise, and image cropping.

Keywords- DWT, Copyright protection, Image processing,Watermarking.

I. INTRODUCTION

Digital watermarking has gained wide acceptance as aviable tool for copyright protection of digital images [1].It involves the insertion of a signal such as a pseudo­noise sequence or a trade mark logo into a digital imageso as to produce the watermarked image. The insertedsignal serves as an identity of the owner and can berecovered from the watermarked image if proof ofownership is required. The watermarked image must berobust to standard image processing procedures and tovarious attacks that may be used to remove or obscure thewatermark. For example, images are often subjected tolossy compressing and may become corrupted withadditive gaussian noise in course of transmission. Thus, awatermarked image is required to be robust against thesetypes of non-malicious attacks.

Attempts to remove watermarks may involve processessuch as cropping and low pass filtering and therefore thewatermarked image should also be robust to these. It isalso a requirement that a good watermarking schemeshould have an invisible watermark so at to lower the risk

Elijah MwangiDept of Electrical & Electronic Engineering

University ofNairobiPO BOX 30197Nairobi, Kenya

elijah.mwangi@uonbLac.ke

ofunauthorised removal or obscuring some desired imagefeatures. Additionally, the security of the watermarkingprocess is also enhanced by the use of secret codes thatare used to embed the watermark.[2]. Some schemes havealso been development to resist geometric attacks such asrotation, scaling, and translation of the watermarkedimage [3].

Since only the watermarked image is released to theauthorised recipients, it is required to have highperceptual quality and to be virtually indistinguishablefrom the original. These two criteria of robustness andperceptual quality are generally inversely related in thathigh robustness requires a watermark signal with moreenergy and this ultimately leads to noticeable distortion inthe watermarked image. Thus, a practical watermarkingscheme must provide a good balance between these tworequirements. The Discrete Wavelet Transform (DWT)domain approach has been reported to give watermarkedimages with high perceptual quality due to theresemblance to the human visual system [4] and hastherefore been employed in our investigations. Ouralgorithm has also been influenced by the recenttechniques in reversible watermarking where arelationship between a quad of pixels is formulated toprovide a novel data hiding strategy [5].

The rest of the paper is organised as follows. Theproposed watermark embedding algorithm is described insection 2. The detection algorithm is presented in section3. The proposed scheme was tested against maliciousattacks and results obtained by computer simulation aregiven in section 4. Lastly, conclusions and suggestionsfor further work are given in section 5.

II. EMBEDDING ALGORITHM

Let I(m,n); 1s m SM, 1s n ~ be the MxN gray-scalecover image which is to be watermarked and B(p,q) 1s pSP; 1s qs: Q, be a PxQ binary image which is to be usedas the watermark. Let the cover image be a square imagei.e, M=N. The embedding of the watermark is illustratedin Fig. 1, and described in the following steps:

978-1-4244-3919-5/09/$25.00 ©2009 IEEE

IEEE AFRICON 2009 23 - 25 September 2009, Nairobi , Kenya

(i) Decompose the image I into l-level DWT to get theAI, HI, VI, and DI quadrants. Where Al is the lowfrequency quadrant, HI the horizontal, VI the vertical,and D1 the diagonal quadrant.

(ii) Partition HI and L1 into non-overlapping blocks ofDWT coefficients of size axa. Where a is even forconvenience.

(iii) Using the secret key, generate the binary matrixK(RxR) where R=M/(2a) . Select position of h, in the axablock.

III. RETRIEVAL ALGORITHM

In the watermark detection and retrieval phase , thesame key that was used in embedding is employed. Thekey helps to identify the watermarked blocks in the DWTtransformed image and also identifies the DWTcoefficient that was perturbed in the imbedding stage.The retrieval process has the following steps and isfurther illustrated in Fig. 2.

(i) Decompose the watermarked image I into l-levelDWT to get the AI, HI , L1, and D 1 quadrants.

(iv) for i=I,2, ...R and forj=I,2 ...Rif K(i,j) ==1, and B(i,j) ==1 embedone I bit ofB in HI as follows:make h; >m '+,1,0"where m'=mean ofaxa matrix, andif is the variance, A, is anarbitrary constant.

(ii) Partition HI and L1 into non-overlapping blocks ofDWT coefficients of size axa.

(iii) Using the secret key, generate the binary matrixK(RxR) where R=M/(2a). select position of h, in the axablock.

(iv) for i=1,2, ...R andj= I,2, ..R

ifK(i,j) ==1, and B(i,j) ==0 embedone 0 bit of B in HI as follows:make h; <tm '-A,CT)

end for loop

if K(i,j) ==1and h, >m' then, B '(i.j) =1else B '(i,j)=0

(v) repeat step (iv) for quadrant VI

(vi) Take inverse DWT and transform back to intensitylevel.

where m'=mean ofaxa matrixelse if K(i,j) =0skip

end for loop

Cover image (v) repeat step (iv) for quadrant VI

VIOl

Al HI

DIVI

(vi) Reconstruct the binary watermark B(PxQ) .

axaBlocks

HIAI

axaBlocks

Fig. I. The watermark embedding process.

BinaryLogo

WATERMARKINSERTION

Secret key Watermarkedimage

axaBlocks

axaBlocks

The embedding algorithm only changes one DWTcoefficient in a pre-determined block in the HI and VIquadrants. The level of watermarking is controlled by avariation of the coefficient ,1,.

RetrievedWatermarked

Fig.2. The watermark retrieval process.

978-1-4244-3919-5/09/$25.00 ©2009 IEEE 2

IEEE AFRICON 2009 23 - 25 September 2009, Nairobi , Kenya

0.8 .----~--~--~--~-----,

~Fig. 6.(a) The watermarked image with Gaussian noise. (b) the retrievedwatermark .

The watermarked images were subjected to variousattacks and the watermark recovery attempted. Theresults are shown below with respect to the 'cameraman'test image. Similar results were obtained with other testimages but not listed here due to space limitation.

A. Gaussian noise attackThe watermarked image was corrupted with additive

zero-mean Gaussian noise and the watermark extracted atvarious levels of input noise variance. At low noiselevels, say for a noise variance of 0.001 and below, thewatermarked was successfully extracted. High noiselevels, say at a variance of 0.1 and above, would requirethe use of an image restoration algorithm to suppress thenoise before watermark retrieval can be attempted. Thereare obvious upper limits at which the watermark wouldbe lost but at the same time it would make the attackedimage less useful to the illegal recipient as recovery of aclean copy would be difficult. The image shown in Fig.6(a), was corrupted with gaussian noise of variance0.005, and the recovered watermark is shown in Fig. 6(b).

0.6

I-Z!:!:! 0.4ou:::tt 0.2oo

..L.

.T.Fig. 4.. (a)The watermarked image at PSNR=38dB .(b)The recovered watermark .

200 400 600 800 1000KEY NUMBER

Fig.5. Watermark detector response to the correct key and variousrandom keys.

..a..

.T.

255PSNR =201og lO { M N }

(l / MN )LL(I(m,n)-Iw(m,n)/m»l nel

(1)The original image is illustrated in Fig.3.(a) and the

binary image that serves as a logo is shown in Fig. 1(b).

Fig. 3.. (a) The 256x256 'Cameraman' test image. (b) The 16x16 binary

watermark .

The effectiveness of the proposed scheme was tested byMATLAB computer simulation using the 256x256 grayscale 8-bit standard images referred to as Lena, Barbara,Mandrill, F16 Airforce jet, and the Cameraman. Thebinary image was a 16xl6 cross-like logo.

In the proposed scheme, the Haar DWT [6] wasemployed and the horizontal and vertical quadrantspartitioned into non-overlapping square blocks of sizes4x4 with the coefficient h; chosen randomly using theentries in the secret key. The Peak-Signal-to-Noise-Ratio(PSNR) as defined in equation 1 was used to measure thelevel of watermarking by correlating the original image tothe watermarked image.

IV. SIMULATION RESULTS

The watermarked image at 38dB is shown in FigA (a)and the extracted binary watermark in Fig. 4(b). It can benoted that the watermarked image has a very highresemblance to the original which indicates that thewatermarking process has little or no statistical effect onthe DWT coefficients. The recovered watermark has acorrelation coefficient of 0.78 to the original.

The effectiveness of the secret key was demonstratedby the response of the watermark detection process to 999randomly generated keys and the correct key. For eachkey, the extracted watermark was correlated to theoriginal watermark. The results shown that only thecorrect key, which is indicated as number 300, could beused to detect the watermark as shown in Fig. 5.

The watermark detection is a simple process that relieson the accuracy of the secret key. The retrieval is a binarydecision that is dependent on the level of coefficientperturbing that was introduced in watermark embedding.It is for this reason that a binary watermark is preferred toa gray-scale type.

978-1-4244-3919-5/09/$25.00 ©2009 IEEE 3

IEEE AFRICON 2009 23 - 25 September 2009, Nairobi , Kenya

B. CroppingThe watermarked image was cropped by cutting off the

top, the bottom, and the sides. The cut-off regions werefilled up with white pixels so as to retain an image of256x256 in size. The watermark was still recovered evenfor heavy clipping levels of down to 50%. a level, atwhich a large part of the image details are lost. The watermarked image that has been cropped to 50% is shown inFig. 7(a), and the extracted water in Fig. 7(b).

Fig.7. (a) The watermarked image. (b) The retrieved watermark.

Fig. 8.. (a) The watermarked image filtered by an l lxl l low passgaussian filter. (a) The extracted binary watermark

[4] E. Elbasi and A.M. Eskicioglu:, "A DWT-based robust semi-blindimage watermarking algorithm using two bands". IS&TISPIE llfh

Symposium on Electronic imaging, Security, Steganography, andWatermarking of Multimedia contents VII conference, San Jose, CA,USA, 15-19 January 2006. pp 1-11.

Fig. 9.. (a) Watermark image (b) Extracted binary watermark

V. CONCLUSION

REFERENCES

This paper has presented a composite blindwatermarking scheme that embeds an invisible binarywatermark image into a gray-level cover image. This isdone without any significant distortion of the originalimage. Experimental results demonstrate that this methodis robust against common image processing operationssuch as low pass filtering, gaussian noise addition,cropping.

However, results obtained with images compressed by afactor greater than 10, show poor recovered of the binarywatermark. Future efforts shall focus on how theproposed algorithm can be used to detect the watermarkin highly compressed IPEG 2000 images.

[2]. Mwangi, E, "A wavelet based image watermarking scheme with aCDMAlHadamard embedding technique". rjh International symposiumon signal processing and its applications. (ISSPA 2007), Sharjah,United Arab Emirates. 12-25th February 2007.

[3] C.Y.Lin, M.wu, J.A. Bloom, and l.J.Cox., "Rotation, scale, andtranslation resilient watermarking for images" . IEEE Transcations onImage Processing Vo1.10, NO.5. pp 767-787. May 2001.

[1] G.C.Langelaar, 1. Setyawan, and RL.Lagendijk, "Watermarkingdigital image and video data: A state of the art review," IEEE signalprocessing magazine No.17, pp. 20-46. September 2000.

'I.L.~

. ...:!It!f"' "I·. :

C. Attack by low pass filteringIn this experiment, the watermarked image was low

pass filtered by a square gaussian low pass filter of agiven size and variance [7]. The variation of the filter sizeand variance with watermark recovery was simulated andresults showed good detection of the watermark forvarious sizes from 3x3, 4x4, up to l lx l l and forvariances between 0.5 and unity. The low pass filteredwatermarked signal using a llxll filter and 0.5 varianceis shown in Fig. 6(a). The extracted binary mark isillustrated in Fig. 6(b).

D. JPEG Compression attackA lossy compression of the watermarked image with a

quality factor Q=80 that gave a compression ratio of 5,was performed and the watermark successively detected.Other compression ratios were employed and successiverates of watermark retrieval noted. The image shown inFig. 8 is IPEG compressed at a factor of 5.3, and theextracted watermark is also shown in the same figure.

[5] S. Weng, Y. Zhao, 1.S. Pan, and R Ni: "A novel reversiblewatermarking based on an integer transform." International Conferenceon Image processing (ICIP 2007). pp III-241 to III-244. 2007.

[6] RC. Gonzalez and RE. Woods: Digital Image Processing. 2nd

Edition. Prentice Hall. New Jersey, U.S.A. 2002.

17] A.K. Jain: Fundamentals of Digital Image Processing. 2nd Edition.Prentice Hall. New Jersey, U.S.A. 1989.

978-1-4244-3919-5/09/$25.00 ©2009 IEEE 4