A Robust Image Watermarking Scheme Invariant to Rotation, Scaling and Translation Attacks

Felix O. Owalla, Student Member, IEEE and Elijah Mwangi, Member, IEEE

Department of Electrical and Information Engineering University of Nairobi

P.O. Box 30197 – 00100 Nairobi, Kenya.

f.owalla@students.uonbi.ac.ke

Abstract— A robust digital image watermarking scheme that is resistant to both signal processing and geometric attacks is proposed. The watermark is embedded in the Discrete Cosine Transform domain in a spread-spectrum format and Vector Quantization techniques used to compress the image. The recovery of the watermark after rotation, scaling and translation attacks is done by using Harris corner detector-based feature-points to get a Delaunay tessellation which is used to reverse the attacks. In situations where RST attacks lead to formation of substantial dark areas on the image, some reference feature-points are lost and recovered watermark is poor or entirely lost. In our proposed scheme, a procedure of estimating the RST attacks is employed by taking an average of selected triangles in the tessellation. Computer simulation results using MATLAB have been used to confirm the accuracy of our proposed scheme.

Index Terms: Delaunay triangulation, Spread-spectrum and Vector Quantization.

I. INTRODUCTION Watermarking has been used extensively to protect digital

media from illegal copying and reproduction. However, this is equally threatened by attackers who use a variety of attacks to remove or to render the watermark useless. These attacks can be roughly grouped into signal processing attacks or geometric attacks. Geometric attacks are difficult to deal with as they involve displacement of pixels thereby inducing synchronization errors between the original and extracted watermarks during detection process [1].

The watermark should also be imperceptible and should not degrade the quality of the image. In order to attain this, the watermark can be embedded in a domain such as the Discrete Cosine Transform (DCT), using the mid frequencies which do not contain visually important features of the image and hence largely unaffected by filtering and noise attacks. The DCT is sufficiently robust to signal processing attacks but very fragile to geometric attacks [1], [2]. There are a few geometric-distortion focused watermarking schemes; they can be roughly grouped into: moment-based, template-based, invariant domain-based and feature-point-based schemes [2]. Feature-point-based techniques have been reported to offer high resistance to geometric attacks. These feature-point-

based techniques include Harris corner detector and Mexican hat wavelet [3], [4]. The Mexican hat wavelet however gives synchronization errors when the geometric attacks are local and therefore Harris corner detector is preferred. The feature-points obtained using Harris corner detector are then used for Delaunay triangulation and Voronoi tessellation which are used for synchronization or generally define the embedding regions [3], [4]. For the images to be efficiently stored and transmitted Vector Quantization (VQ) is often used to compress the image.

Some extracted feature points are lost when the attacks lead to formation of substantial dark regions around the image and hence giving a different Delaunay tessellation. The difference in the tessellation leads to synchronization errors as it is more difficult to estimate the rotation, scaling and translation (RST) attack. Our proposed scheme uses averaging of triangles to estimate the exact nature of attack on the image in order to restore it to the position of the original image. This ensures that the watermark can still be extracted even after losing a number of reference points.

This paper is organized as follows. Section (II) describes the extraction of feature points. Section (III) describes Delaunay triangulation. Section (IV) describes the watermark embedding process, the watermark recovery procedure and the proposed scheme presented. Section (V) presents the experimental results and the conclusion given in Section (VI).

II. FEATURE-POINTS EXTRACTION

After reconstructing the image from the watermarked codebook of the image, the feature-points which will be used in synchronizing the attacked image are obtained [4]. Robust feature-points are obtained using Harris corner detector. This is done by obtaining feature-points and then rotating the image by various angles and repeating the above process, the features which are extracted for all the angles are then taken as the robust feature-points.

Harris corner detector is based on local auto-correlation function of a signal. Local auto-correlation functions measure

978-1-4673-0784-0/12/$31.00 ©2012 IEEE 379

the local changes of the signal with patches amount in different directions.

Given a shift (Δx,Δy) and a point (x,y), thfunction is defined as

, Δ Δ , ΔΔ . Where matrix C(x,y) is given by , ∑ , ∑ ,∑ , , ∑Where w is a Gaussian window of say

pixels centered at (x,y) and Ix(x,y) and Iderivatives of the image function I(x,y) irespectively.

The matrix C(x,y) captures the intensitylocal neighbourhood. The eigenvalues of thto obtain the local auto-correlation function image region. When both the eigenvaluwindowed region is flat, when one eigenvaluother one low the region is an edge and whthen the region is a corner.

The feature-points are then used in obtriangulation. The triangles are employed the image during watermark extraction [3], [

III. DELAUNAY TRIANGUL

The Delaunay tessellation of the featurefrom an image is obtained by determinneighbours [3]. In order to get the Delaunthe feature points, a Voronoi diagram of the first.

Voronoi tessellation is constructed by deconsisting of all locations in the space whichpoint pi than to any other point on the imageseparating two adjacent Voronoi regions Di points pi and pj is a perpendicular bisector of

Fig. 1 illustrates the Voronoi and Delaunaa set of points.

(a) (b)

Fig 1 (a) Voronoi tessellation of a set of points antessellation.

To construct the Delaunay tesse

neighrbouring points, whose cells in the Voshare an edge are joined. The Delaunay tessof feature points of images are illustrated in f

shifted by a small

he auto-correlation

1

, ,, .

(2) 3-by-3 or 5-by-5

Iy(x,y) are partial in x and y axes

y structure of the he matrix are used

of the windowed ues are low, the ue is high and the hen both are high

btaining Delaunay in synchronizing

[5].

LATION e-points extracted ning the nearest

nay tessellation of image is obtained

fining the area Di h are closest to a e. The boundaries and Dj containing f the two points. ay tessellations of

)

nd their (b) Delaunay

ellation nearest oronoi tessellation sellations of a set fig. 2.

Delaunay tessellation is a dual ofemploys the concept of an empty circumcircle criterion is such thaforming a Delaunay triangle lie onpoint should lie inside the circircumcircle criterion gives Delamaximizing the minimum angleminimizing the size of the circumciralso has the property of being uniqucircumcircle, which is four points lay

Fig 2 Delaunay triangulations of the feat

Lena test image

IV. PROPOSEDIn this paper an algorithm which

Spread-spectrum in the watermark incorporates Delaunay triangulation The embedding procedure can be sumLBG algorithm is used to design anthe encoded image, the codebook is by-8 pixel blocks [7], - [9] and the block obtained. The DCT coefficienthen embedded into the coefficienspread-spectrum format.

The watermark extraction procefollows. Delaunay triangulation of image is used to resynchronize the the original image. After resynchroobtained and the DCT coefficiwatermark extracted from the embed

A. Embedding Proced

1. Step 1. The LBG algorithm tothe image with a codeword size

2. Step 2. Decompose the image codebook and compute a bcodebook and select the cofrequency sub-band.

3. Step 3. Generate a pseudo-ranthe value of β which controls th

4. Step 4. Embed the watermarkfrequency sub-band using thsequence to determine the watermark is embedded as follo

, ,

f Voronoi tessellation and circumcircle. The empty

at for any three points n the circumcircle and no rcumcircle. The empty

aunay the properties of e of the triangles and rcle. Delaunay tessellation ue except in cases of dual ying on a circumcircle.

ture points for Cameraman and es

METHOD h combines VQ, DCT and

embedding process and in the extraction process.

mmarized as follows. The nd obtain the codebook of again decomposed into 8-DCT coefficients of each nts of the watermarks are

nts of the codebook in a

ss can be summarized as the feature points of the image to the position of

nization, the codebook is ients obtained and the dded positions.

dure o construct a codebook of e of 8-by-8. into 8-by-8 blocks of the lock-based DCT of the oefficients of the mid-

ndom sequence and select he embedding strength. k into the selected mid-he pseudo-random (PN)

coefficients where the ows: ,

(3)

380

Where, I(u,v), Iw(u,v) and W(u,v) arwatermarked and watermark DCT coecodebook respectively. 5. Step 5. Compute the inverse DCT of

codebook and then perform VQ decodwatermarked image.

The value of β between 0.1 and 0.5 was foutrade of between robustness and visual qualthe pixel texture of the image [10]. It was generally smooth pixel images a higher valused with less visual quality distortion thanimages. B. Watermark Recovery Procedure

The process of watermark recovery frogiven in this subsection as follows:

1. Step 1. Obtain the feature-points of the

Harris corner detector then obtaintriangulation of the feature-points obtai

2. Step 2. Select several target triangles aprobe triangles from the Delaunay ofattacked image respectively.

3. Step 3. Using the triangles estimatedistortion that each probe triangle hcomparing the orientation angles, sizesthe corresponding target triangle.

4. Step 4. Obtain the average of the distoby each triangle and use the averagefactor (RF), scaling factor (SF) and (TF) which are then used to restore original form.

5. Step 5. Obtain the codebook of theimage and compute its DCT.

6. Step 6. Select the mid-band frequencisame PN sequence used in the embeddifor the presence or absence of the embusing correlating the coefficients ofrequencies and the PN sequence.

7. Step 7. Reconstruct the watermark uswatermark bits.

V. EXPERIMENTAL RESTo evaluate the performance of our p

several experiments were performed to detperceptibility and robustness of the embedde

A. Visual Perceptibility The watermarked images were asse

distortion and then Peak Signal-to-Noise Rto determine quality degradation as a result watermark. For example, for the cameramanis no visible visual degradation in quality at 34.9 dB. Similar observations were noteimages.

re the original, efficients of the

the watermarked ding to obtain the

und to give a fair lity depending on observed that the lue of β could be n the rough pixel

om the image is

e image using the n the Delaunay ined. and corresponding f the original and

e the amount of as undergone by s and positions to

ortions undergone es as the rotation translation factor the image to its

e resynchronized

ies and using the ing process check

bedded watermark of the mid-band

sing the extracted

SULTS proposed scheme, ermine the visual

ed watermark.

essed for visual atio (PSNR) used of the embedded

n test image there a PSNR value of

ed for other test

B. Robustness to AttaThe watermarked image was s

attacks, including signal processindistortion attacks. The watermark restorations where necessary and thecomputer using a Normalized cross-

The first attack to be simulatedSeveral rotation attack angles werreversal was done for each angextracted. Fig. 3 illustrates the proban RF of 790. The embedded waterthat was recovered after restoring toriginal position are shown in fig. 4has an NC of 0.99.

(a)

Fig. 3 The (a) probe triangle and (b) targetdistortion RF = 7

(a)

Fig. 4 Watermark embedded in cameraman i(b) extracted message after resto

The second attack to be simulateThe image was scaled from 25% image size. At 25% of the original watermark was unsatisfactory, wh50% the watermark is recoverestoration. Fig. 5 shows the tessellthe watermark recovered from the at0.5.

(a) (b)

Fig. 5 (a) Delaunay triangulation, (b) the probmessage at SF =

Translation attacks for various

and the watermark was recovered sa48 pixels after which it was no

Original Message

acks subjected to a variety of ng attacks and geometric

was then extracted after e quality of the watermark correlation (NC) factor.

d was the rotation attack. re simulated and then a gle and the watermark

be and target triangle with rmark and the watermark the attacked image to its 4. The recovered message

(b)

t triangle after a huge rotation 790.

(b)

image (a) original message and oring rotation attack

ed was the scaling attack. to 125% of the original image size the recovered

hile scaling levels above ered successfully after lation, probe triangle and ttacked image at an SF of

(c)

be triangle and (c) the recovered 0.5

TF were also simulated atisfactorily up to a TF of

ot recognizable. Fig. 6

Recovered Message

381

illustrates the Delaunay tessellation and protranslating the image by a TF of (25,0)message is shown in fig. 6 (c).

(a) (b)

Fig. 6 (a) Delaunay triangulation and (b) the probe trian(c) Recovered message for the translation

Finally combinations of two or all the t

are simulated and the embedded watermarkrestoring the image to its original position.an image subjected to a combination of RSretrieved watermark after a combination rotation, rotation and translation and RST tuple factor of 120, 0.75 and (20,5).

(a) (b)

Fig.7(a) Probe triangle after a combination RST attamessages for a combination of (b) scaling and rotatio

translation and (d) rotation, scaling and transla

Computer simulation experiments werwith other common test images such as BLena and Baboon among others. These experobustness of the watermark to signal procesas JPEG compression and histogram equaliza

TABLE I.

NC OF VARIOUS ATTACKS ON VARIOUSPROPOSED SCHEME VS TANG

Cameraman Lena Ours Tang Ours T

JPEG (30%) 0.98 0.96 0.99 0JPEG (50%) 0.99 0.98 0.99 0

Avg. filt. 0.97 0.97 0.97 0Unsharp filt. 0.99 0.98 0.98 0Histo. equal. 0.97 0.97 0.99 0Cropping 0.1 0.89 0.72 0.93 0Rotation (79o) 0.99 0.92 0.99 0Scaling (0.5) 0.94 0.87 0.91 0

Translation(25,0) 0.99 0.86 0.97 0

RST 12o,0.75,(20,5) 0.91 0.76 0.89 0

obe triangle after ). The recovered

(c)

ngle TF = 25,0 pixels n attack

three RST attacks ks extracted after Fig. 7 illustrates T attacks and the of scaling and with a 3-element

(c) (d)

acks and recovered on, (c) rotation and ation attacks

re also conducted Barbara, Peppers, eriments tested the ssing attacks such ation.

S IMAGES (OUR G’S [3]) a Baboon Tang Ours Tang 0.98 0.98 0.98 0.97 0.97 0.98 0.98 0.99 0.96 0.97 0.96 0.94 0.95 0.99 0.94 0.71 0.95 0.77 0.89 0.97 0.94 0.79 0.89 0.81

0.76 0.97 0.79

0.74 0.85 0.71

In all situations the waterrecovered with a high NC factor andshown in table 1. In addition the teto RST attacks and our proposed schthe watermark. This was also succeshown in table 1.

VI. CONCLU

In this paper we have demonstrabe recovered after RST attacks on Delaunay tessellation techniques. Iattacks lead to formation of large dascheme of averaging has been shrecovering the watermark at high NCbeen tested with success on varMATLAB simulation platform.

REFERENCE

[1] S. Katzenbeisser, and F.A.P. PetiTechniques for Steganography and DUSA: Artech House Inc. 2000.

[2] R. Gonzalez, R.E. Woods, S.LProcessing, 3rd Edition, New Delhi, InLearning Pvt. Ltd, 2008.

[3] C.-H. Tang and H.-M. Hang. “A FDigital Image Watermarking Scheme,” IVol. 51, No. 4, pp. 950-959, 2003.

[4] X. Qi and J. Qi. “A feature-watermarking scheme,” Proc. 7th IASTImage Process. (SIP ’05), Honolulu, H2005.

[5] C. Harris and M. Stephens. “A detector,” Proc. 4th Alvey Vision Confere

[6] P. Bas, J.-M. Chassery, and B. MacWatermarking Using Feature Point,” IEVol. 11, No. 9, pp. 1014-1028, 2002. [7] A. K. Pal, S. Das, G.P. Biswas and DCT Scheme for Grayscale Image WComput. Applicat., Vol.2 – No.3, pp 54 –

[8] H.C. Huang, S.C. Chu and Watermarking Techniques,” Journal opp.37-50, July 2006.

[9] Y. Linde, A. Buzo, and R.M. Grayquantizer design,” IEEE Trans. Commun1980.

[10]X. Li, B. Yang and T. ZeWatermarking Based on Adaptive PredPixel Selection”, IEEE Trans. Image Pr3524-3533, 2011.

rmark was successfully d the numerical results are est images were subjected heme employed to recover essful and the results are

USION ated that a watermark can

an image by employing In situations where RST ark regions, our proposed hown to be effective in C factors. The scheme has rious test images on a

ES

itcolas, Information Hiding Digital Watermarking, MA,

L. Eddins, Digital Image ndia: Prentice Hall of India

Feature-Based Robust Image IEEE Trans. Signal Process.,

-point-based RST resistant TED Int. Conf. Signal and

Hawaii, U.S.A., pp. 469-474,

combined corner and edge ence, pp. 147-151, 1988.

cq, “Geometrically Invariant EEE Trans. Image Process.,

S. Mukhopadhyay, “A VQ-Watermarking,” Int. Journal – 59, May 2010.

Y.H. Huang, “VQ-Based of Comput., Vol.17, No.2,

y, “An algorithm for vector n., Vol. 28, No. 1, pp. 84-95,

eng, “Efficient Reversible diction-Error Expansion and rocess., Vol. 20, No. 12, pp.

382