Sie sind auf Seite 1von 18

Expert Systems with Applications 38 (2011) 20812098

Contents lists available at ScienceDirect

Expert Systems with Applications


journal homepage: www.elsevier.com/locate/eswa

A novel color image watermarking scheme in nonsampled contourlet-domain q


Pan-Pan Niu a, Xiang-Yang Wang b,, Yi-Ping Yang b, Ming-Yu Lu a
a b

School of Information Science & Technology, Dalian Maritime University, Dalian 116026, China School of Computer and Information Technology, Liaoning Normal University, Dalian 116029, China

a r t i c l e

i n f o

a b s t r a c t
Geometric distortion is known as one of the most difcult attacks to resist, for it can desynchronize the location of the watermark and hence causes incorrect watermark detection. It is a challenging work to design a robust color image watermarking scheme against geometric distortions. Based on the support vector regression (SVR) and nonsubsampled contourlet transform (NSCT), we propose a new color image watermarking algorithm with good visual quality and reasonable resistance toward geometric distortions in this paper. Firstly, the geometrically invariant space is constructed by using color image normalization, and a signicant region is obtained from the normalized color image by utilizing the invariant centroid theory. Then, the NSCT is performed on the green channel of the signicant region. Finally, the digital watermark is embedded into host color image by modifying the low frequency NSCT coefcients, in which the HVS masking is used to control the watermark embedding strength. In watermark detection, according to the high correlation among different channels of the color image, the digital watermark can be recovered by using SVR technique. Experimental results show that the proposed color image watermarking is not only invisible and robust against common image processing operations such as ltering, noise adding, and JPEG compression etc., but also robust against the geometrical distortions. 2010 Elsevier Ltd. All rights reserved.

Keywords: Image watermarking Nonsubsampled contourlet transform (NSCT) Color image normalization HVS masking

1. Introduction With the rapid growth and widespread use of network distributions of digital media content, there is an urgent need for protecting the copyright of digital content against piracy and malicious manipulation. Digital watermarking has been proposed as a possible and efcient answer to these concerns. While the most prominent application of watermarking is copyright protection, others including ngerprinting, broadcast monitoring, digital media authentication, and copy protection are important research areas (Lian, Kanellopoulos, & Ruffo, 2009). For different purposes, digital watermarking has been branched into two classications: robust watermarking technique and fragile watermarking technique. Ro-

q This work was supported by the National Natural Science Foundation of China under Grant Nos. 60773031 & 60873222, the Open Foundation of State Key Laboratory of Networking and Switching Technology of China under Grant No. SKLNST-2008-1-01, the Open Foundation of Network and Data Security Key Laboratory of Sichuan Province, the Open Foundation of State Key Laboratory for Novel Software Technology of China under Grant No. A200702, the Open Foundation of Key Laboratory of Modern Acoustics Nanjing University under Grant No. 0802, and Liaoning Research Project for Institutions of Higher Education of China under Grant No. L2010230. Corresponding author at: School of Computer and Information Technology, Liaoning Normal University, Dalian 116029, China. Tel.: +86 0411 85992415; fax: +86 0411 85992323. E-mail address: wxy37@126.com (X.-Y. Wang).

bust digital watermarking is used to protect ownership of the digital media. In contrast, the purpose of fragile watermarking technique is digital media authentication, that is, to ensure the integrity of the digital media. In recent years, there is an unprecedented development in the robust image watermarking eld. On the other hand, attacks against image watermarking systems have become more sophisticated (Zheng, Liu, Zhao, & El Saddik, 2007). In general, these attacks on watermarking systems can be categorized into common image processing operations and geometrical distortions. Watermark robustness to common image processing operations has been addressed extensively, such as lossy compression, noise addition and median ltering etc. However, many of the existing schemes are fragile to geometrical distortions, such as rotation, scaling and translation (RST). Even a slight geometric distortion may signicantly inuence the extraction of watermark because the synchronization of the watermark is destroyed. Nowadays, several approaches that counterattack geometric distortions have been developed. These schemes can be roughly divided into invariant transform, template insertion and feature-based algorithms (Dong, Jovan, Yongyi, Yang, & Franck, 2005; Zheng et al., 2007; Zheng, Wang, & Zhao, 2009). Invariant transform: The most obvious way to achieve resilience against geometric distortions is to use an invariant transform. In Zheng et al. (2007), Guo, Liu, and Liu (2007), Xin, Liao, and Pawlak (2007), Zhang, Qian, Xiao, and Ji (2007) and Xiang,

0957-4174/$ - see front matter 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.07.147

2082

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

Kim, and Huang (2008), the watermark is embedded in an afne invariant domain by using FourierMellin transform, generalized Radon transform, pseudo-Zernike moments, geometric moments, and histogram shape respectively. In Zhang, Cheng, Qiu, and Cheng (2008), an algorithm of leveled watermarking is proposed. This algorithm is based on the curvelet transform and does not involve the original image at the detector end or exhaustive direction search. Watermark bits with different robustness are embedded in each scale and one bit is carried by one wedge. The coefcients selection for watermarking is based on the conception that the coefcient energy is proportional to its directional sensitivity. Tai, Yeh, and Chang (2009) presented a reversible data hiding scheme based on histogram modication. They exploit a binary tree structure to solve the problem of communicating pairs of peak points. Distribution of pixel differences is used to achieve large hiding capacity while keeping the distortion low. They also adopt a histogram shifting technique to prevent overow and underow. Despite that they are robust against global afne transformations, those techniques involving invariant domain suffer from implementation issues and are vulnerable to mixed attacks. Template insertion: Another solution to cope with geometric distortions is to identify the transformation by retrieving articially embedded references. Liu, Zheng, and Zhao (2007) presents an image rectication scheme that can be used by any image watermarking algorithm to provide robustness against rotation, scaling and translation (RST). In the watermarking, a small block is cut from the log-polar mapping (LPM) domain as a matching template, and a new ltering method is proposed to compute the cross-correlation between this template and the magnitude of the LPM of the image having undergone RST transformations to detect the rotation and scaling parameters. Zhang, Li, and Wang (2008) designs look-up table (LUT) according to the distortion of LUT embedding. A new practical reduced-distortion LUT design method is developed for robust data hiding. A Gaussian mixture model and a related expectation-maximization algorithm based method are employed to model the statistical distribution of the host image. The statistical model is used to select signicant wavelet coefcients of the host image for data hiding. Boato and Conotter (2009) present an innovative and exible tool suitable to assess the robustness of digital watermarking techniques, by introducing a novel metric based on the perceptual quality evaluation for unmarked images. Genetic algorithms (GA) perform the search of optimal parameters to be assigned to each image processing operator, as well as the order they need to be applied in, to remove the watermark from the content while keeping the perceptual quality of the resulting image as high as possible. However, this kind of approach can be tampered with by the malicious attack. As for random bending attacks, the template-based methods will be incompetent to estimate the attack parameters. Feature-based: The last category is based on media features. Its basic idea is that, by binding the watermark with the geometrically invariant image features (Local Feature Region, LFR), the watermark detection can be done without synchronization error. Lee, Lee, and Lee (2007) propose a geometrically invariant watermarking method that uses circular Hough transform for watermark synchronization. Through circular Hough transform, the circular features are extracted that are invariant to geometric distortions. Based on HarrisLaplace detector and scale-space theory, Wang, Wu, and Niu (2007) propose a feature-based digital image watermarking scheme in DFT domain. In Li and Guo (2009), present a novel robust image watermarking scheme for resisting geometric distortions. Watermark synchronization is rst achieved by local invariant regions which can be generated using scale normalization and image feature points. The water-

mark is embedded into all the local regions repeatedly in spatial domain. Deng, Gao, Li, and Tao (2009) give a content-based watermarking scheme that combines the invariant feature extraction with watermark embedding by using Tchebichef moments. Pham, Miyaki, Yamaski, and Aizama (2008) present a robust object-based watermarking algorithm using the local image feature in conjunction with a data embedding method based on DCT, and the digital watermark is embedded in the DCT domain of randomly generated blocks in the selected object region. Gao, Deng, and Li (2010) propose a new image watermarking scheme by incorporating the advantages of the afne invariant point detector and the orientation alignment seamlessly. The afne invariant point detector is adopted to extract Afne Covariant Regions (ACRs). The graph theoretical clustering algorithm is then employed to select a set of nonoverlapped ACRs for watermark embedding. Singhal, Lee, and Kim (2009) propose a robust image watermarking algorithm using local Zernike moments, which are computed over circular patches around feature points. The proposed algorithm locally computes Zernike moments and modies them to embed watermarks, achieving robustness against cropping and local geometric attacks. Moreover, to deal with scaling attacks, the proposed algorithm extracts salient region parameters, which consist of an invariant centroid and a salient scale, and transmits them to the decoder. The parameters are used at the decoder to normalize a suspect image and detect watermarks. In Li, Guo, and Pan (2010), a new robust image watermarking scheme is presented by combining scale-space feature point based watermark synchronization and NSCT based watermark embedding. Watermark synchronization is achieved based on the local circular regions, which can be generated using the scale-invariant feature transform (SIFT). In the encoder, the watermark is embedded into the NSCT coefcients in a content-based and rotation-invariant manner by oddeven quantization. In the decoder, the watermark can be extracted directly from the local regions using the proposed coefcient property detector (CPD). It is not difcult to see that the feature-based approaches are better than others in terms of robustness. However, some drawbacks indwelled in current feature-based schemes restrict the performance of watermarking system. First, the feature point extraction is sensitive to image modication. Second, the computational complexity in calculating the features of an image before watermark detection is added. Third, the volume of watermark data is lesser. Generally speaking, the above-mentioned research efforts concentrated on developing the watermarking schemes for gray images, and the watermarking schemes for color images are very few. While, in ubiquitous multimedia applications, color images are basic components of multimedia systems, such as video systems based on current video compression standards, MPEG-1, MPEG-2, MPEG-4, and further MPEGs. Hence, it is crucial to rst develop effective watermarking techniques for color images. In Peng and Liu (2008), a new semi-fragile watermarking scheme for color image authentication is proposed based on spatiotemporal chaos and singular value decomposition (SVD). The watermark is embedded into the singular values of the blocks within wavelet subband. In order to enhance the security, spatiotemporal chaos is employed to select the embedding positions for each watermark bit as well as for watermark encryption. Tsui, Zhang, and Androutsos (2008) encodes the L*a*b* components of color images and watermarks are embedded as vectors by modifying the Spatiochromatic Discrete Fourier Transform (SCDFT) coefcients and using the Quaternion Fourier Transform (QFT). However, it was found that even very small geometric distortions, such as RST attacks, could reduce the detectors ability of detecting watermarks. Arash and Shohreh (2008) proposed a novel method to utilize the Principal Component Analysis (PCA) in the neighborhoods of an image in

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

2083

order to extract the corresponding eigenimages. These eigenimages exhibit high levels of energy compaction and thus are appropriate for such operations as compression and watermarking. However, the algorithm is not robust to cropping. In Lee and Kwon (2008), a color image watermarking algorithm based on Direct Sequence-Code Division Multiple Access (DS-CDMA) and Hadamard kernel has been proposed. In Liu (2010), a wavelet-based watermarking scheme for color images is proposed. The watermarking scheme is based on the design of a color visual model that is the modication of a perceptual model used in the image coding of gray scale images. The model is to estimate the noise detection threshold of each wavelet coefcient in luminance and chrominance components of color images in order to satisfy transparency and robustness required by the color image watermarking technique. Unfortunately these approaches are non-invariant to image rotation and translation attacks. Zheng and Feng (2009) proposed a new watermarking scheme based on the multi-channel image watermarking framework, which generates a watermarking template from one of image channels data, and then embeds this watermarking template into another image channel. In Lin, Shie, and Guo (2010), an improved DCT-based image watermarking technique has been introduced. The watermark is embedded into a digital image, based on the concept of mathematical remainder, by modifying the low-frequency coefcients in DCT frequency domain. Fu and Shen (2008) presented a novel oblivious color image watermarking scheme based on linear discriminant analysis (LDA). The watermark accompanied with a reference is embedded into the RGB channels of color images. By applying the embedded reference watermark, a linear discriminant matrix is obtained. The watermark can be correctly extracted under several different attacks. Behnia, Teshnehlab, and Ayubi (2010) proposed a new watermarking scheme for color image based on a family of the pair-coupled maps. Pair-coupled maps are employed to improve the security of watermarked image, and to encrypt the embedding position of the host image. Another map is also used to determine the pixel bit of host image for the watermark embedding. The purpose of this algorithm is to improve the shortcoming of watermarking such as small key space and low security. These algorithms are robust to image rotation and cropping, but they are not robust to translation and scaling distortion in general. Xing and Tan (2007) gives a color watermarking algorithm not only robust to waveform attacks due to the use of block-SVD, but also robust to cropping because of the watermark Arnold transformation preprocessing. But the scheme only survives very small angle rotation. In Xing and Tan (2010), the authors add a step before watermark extract, that is, calculating phase correlation in image log-polar mapping (LPM) domain to estimate the geometrical distortion parameters to resynchronize the watermark position destroyed by geometrical distortion. Therefore, the scheme is robust to all most geometric distortions, such as rotation, scaling and cropping, but it is very little under translation attacks. In order to effectively resolve the problem of resisting geometric distortions, the support vector machine (SVM) theory is introduced to the image watermarking domain. In scheme (Wu & Xie, 2006), in order to obtain the rotation, scaling and translation (RST) parameters, the SVM are utilized to learn image geometric pattern represented by six combined low order image moments. The watermark extraction is carried out after watermarked image has been synchronized without original image. Wang, Xu, and Yang (2009) propose a robust image watermarking detection algorithm against geometric distortions, in which the steady pseudo-Zernike moments and Krawtchouk moments are utilized. Tsai and Sun (2007) propose a novel watermarking technique called SVM-based color image watermarking (SCIW) for the authentication of color images. The SCIW method utilizes the set of training patterns to train the SVM and then applies the trained SVM to classify a set

of testing patterns. Following the results produced by the classier, the SCIW method retrieves the hidden watermark without the original image during watermark extraction. Li, Ling, and Lu (2007) introduce a novel semi-fragile watermarking scheme based on SVM. This scheme rst gives a denition of wavelet coefcient direction tree, then the relation model between the root node and its offspring nodes is established using SVM, and further watermark is embedded and extracted based on this relation model. Based on a large number of theory analyses and experimental results, we can easily come to the conclusion that it is possible to resist geometric distortions by utilizing the advanced SVM, but the current SVM-based image watermarking have shortcomings as follows: (i) They are not very robust against some geometric distortions, such as cropping, mixed attacks etc; (ii) In watermark detection procedure, the original watermark signal is needed, so it is unfavorable to practical application. In this paper, a new color image watermarking algorithm with good visual quality and reasonable resistance toward geometric distortions is proposed, in which the SVR and NSCT are utilized. Firstly, the geometrically invariant space is constructed by using color image normalization, and a signicant region is obtained from the normalized color image by utilizing the invariant centroid theory. Then, the NSCT is performed on the green channel of the signicant region. Finally, the digital watermark is embedded into host color image by modifying the low-frequency NSCT coefcients, in which the HVS masking is used to control the watermark embedding strength. In watermark detection, according to the high correlation among different channels of the color image, the digital watermark can be recovered by using SVR technique. The rest of this paper is organized as follows. Section 2 presents the basic theory about NSCT. In Section 3, the SVR technique is described. Section 4 contains the description of our watermark embedding procedure. Section 5 covers the details of the watermark detection procedure. Simulation results in Section 6 will show the performance of our scheme. Finally, Section 7 concludes this presentation. 2. An introduction to nonsubsampled contourlet transform (NSCT) Fig. 1(a) displays an overview of the nonsubsampled contourlet transform (NSCT). The structure consists in a bank of lters that splits the 2-D frequency plane into the subbands illustrated in Fig. 1(b). The NSCT can be divided into two shift-invariant parts: (1) a nonsubsampled pyramid structure that ensures the multiscale property and (2) a nonsubsampled directional lter bank (DFB) structure that gives directionality. 2.1. Nonsubsampled pyramid (NSP) The multiscale property of the NSCT is obtained from a shiftinvariant ltering structure that achieves a subband decomposition similar to that of the Laplacian pyramid. This is achieved by using two-channel nonsubsampled 2-D lter banks. Fig. 2 illustrates the nonsubsampled pyramid (NSP) decomposition with J = 3 stages. Such expansion is conceptually similar to the onedimensional (1-D) Nonsubsampled Wavelet Transform (NSWT) computed with the trous algorithm (Cunha, Zhou, & Do, 2006) and has J + 1 redundancy, where J denotes the number of decomposition stages. The ideal passband support of the lowpass lter at the jth stage is the region [(p/2j), (p/2j)]2. Accordingly, the ideal support of the equivalent highpass lter is the complement of the lowpass, i.e., the region [(p/2j1), (p /2j1)]2n[(p/2j), (p/ 2j)]2. The lters for subsequent stages are obtained by upsampling

2084

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

Lowpass subband Highpass Subband Bandpass Directional Subbands Highpass Subband Bandpass Directional Subbands

( , )

Image

( , )
(b)

(a)

Fig. 1. Nonsubsampled contourlet transform: (a) Nonsubsampled lter bank (NSFB) structure that implements the NSCT, (b) Idealized frequency partitioning obtained with the NSCT structure.

Fig. 2. The NSP is a 2-D multiresolution expansion similar to the 1-D NSWT: (a) Three-stage pyramid decomposition. The lighter gray regions denote the aliasing caused by upsampling, (b) Subbands on the 2-D frequency plane.

the lter of the rst stage. This gives the multiscale property without the need for additional lter design. The proposed structure is thus different from the separable NSWT. In particular, one bandpass image is produced at each stage resulting in J + 1 redundancy. By contrast, the NSWT produces three directional images at each stage, resulting in 3J + 1 redundancy. The advantage of NSP is that it is general and as a result, better lters can be obtained. 2.2. Nonsubsampled directional lter bank (NSDFB) The directional lter bank of Cunha et al. (2006) is constructed by combining critically-sampled two-channel fan lter banks and resampling operations. The result is a tree-structured lter bank that splits the 2-D frequency plane into directional wedges. A shift-invariant directional expansion is obtained with a nonsubsampled DFB (NSDFB). The NSDFB is constructed by eliminating the downsamplers and upsamplers in the DFB. This is done by switching off the downsamplers/upsamplers in each two-channel lter bank in the DFB tree structure and upsampling the lters accordingly. This results in a tree composed of two-channel Nonsubsampled Filter Banks (NSFB). Fig. 3 illustrates a four channel decomposition.

Note that in the second level, the upsampled fan lters Ui (zQ), i = 0, 1 have checker-board frequency support, and when combined with the lters in the rst level give the four directional frequency decomposition shown in Fig. 3. The synthesis lter bank is obtained similarly. Just like the critically-sampled directional lter bank, all lter banks in the nonsubsampled directional lter bank tree structure are obtained from a single NSFB with fan lters. Moreover, each lter bank in the NSDFB tree has the same computational complexity as that of the buildingblock NSFB. 2.3. Combining the nonsubsampled pyramid and nonsubsampled directional lter bank in the NSCT The NSCT is constructed by combining the NSP and the NSDFB as shown in Fig. 1(a). In constructing the NSCT, care must be taken when applying the directional lters to the coarser scales of the pyramid. Due to the tree-structure nature of the NSDFB, the directional response at the lower and upper frequencies suffers from aliasing which can be a problem in the upper stages of the pyramid. This is illustrated in Fig. 4(a), where the passband region of the directional lter is labeled as Good or Bad. Thus, we see that

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

2085

Fig. 3. Four-channel nonsubsampled directional lter bank (NSDFB) constructed with two-channel fan lter banks: (a) Filtering structure, (b) Corresponding frequency decomposition.

Fig. 4. Need for upsampling in the NSCT: (a) With no upsampling, the highpass at higher scales will be ltered by the portion of the directional lter that has bad response, (b) Upsampling ensures that ltering is done in the good region.

for coarser scales, the highpass channel in effect is ltered with the bad portion of the directional lter passband. This results in severe aliasing and in some observed cases a considerable loss of directional resolution. Fig. 5 gives experimental results of applying the NSCT decomposition on Zoneplate image, yielding one lowpass subband and eight highpass subbands. Here, NSCT is performed with 3 scale and 8 directions in the nest subband. 3. Support vector regression (SVR) Support vector machine (SVM) is a universal classication algorithm proposed by Vapnik (1995) in the middle of 1990s, it is thought of a new innovation of learning machine, which uses the statistical learning theory. The basic theory of SVM can be depicted by a typical twodimensional case shown in Fig. 6. In Fig. 6,  and j denote two categories of samples. H is the separating hyperplane, H1 and H2 are parallel to H (they have the same normal) and no training points fall between them; the margin of a separating hyperplane is dened as H1 + H2. The optimal separating hyperplane what you call not only can separate the two categories of samples exactly (the ratio of training errors is 0), but also has the maximal margin. Thus the problem of optimal separating hyperplane can be transformed a constraints problem. Support vector regression (SVR) is an application of Support vector machine on regression learning. Next, we will briey introduce the related theories about SVR.

For the training sets:

fx1 ; y1 ; x2 ; y2 ; . . . ; xn ; yn g
x 2 Rn, y 2 R to get the relation between the input xi and output yi, it can seeks a optimal regression function f(x) by SVR training, so that the difference between the output value and the corresponding objective value of every input samples is not more than error e. For the linear situation, the form of function is: f(x) = x x + b, x 2 x, b 2 R. In order to get an optimal regression function, it needs a minimum x, then the above problem can be described as an optimize problem:

min s:t:

1 kxk2 ; 2  yi x xi b 6 e;

x xi b yi 6 e;

i 1; 2; . . . ; n

Considering the existence of regression error is permitted, so the positive slack variables ni P 0 and n i P 0 are introduced, and the above formula can be described as follows:

min s:t:

n X 1 kxk2 C ni n i ; 2 i 1  yi x xi b 6 e ni ;

x xi b yi 6 e ni ;

i 1; . . . ; n

Then, the optimized problem is equal to the following problem according to the Wolfe dual principle:

2086

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

Fig. 5. The NSCT decomposition on Zoneplate image: (a) Origin Zoneplate, (b) The lowpass subband in NSCT domain, (c)(j) The highass subbands at 8 directions in NSCT domain.

A generic separating hyperplane

H1 H

Maximum margin

H2

Support vectors

Class I Class II

The

optimal separating hyperplane


Fig. 6. The basic theory of SVM (Linear separation).

max xa; a e

n X i1

a i ai

n X yi a i ai i1

4. Watermark embedding scheme In this paper, a new NSCT domain color image watermarking algorithm with good visual quality and reasonable resistance toward geometric distortions is proposed. Firstly, the geometrically invariant space is constructed by using color image normalization, and a signicant region is obtained from the normalized color image by utilizing the invariant centroid theory. Then, the NSCT is performed on the green channel of the signicant region. Finally, the digital watermark is embedded into host color image by modifying the low-frequency NSCT coefcients, in which the HVS masking is used to control the watermark embedding strength. The watermark embedding procedure is shown in Fig. 7. Let

n   1X ai ai a j aj xi xj 2 i; j 1

s:t:

n X i 1

ai a 0 6 ai ; a i 0; i 6 c

Where i 1; 2; . . . ; n; ai ; a i is Lagrange multiplying factors, the regression function is:

f x w x b

n X i 1

a i ai xi x b

Where ai ; a i are only few non-zeros which corresponding samples are namely called support vectors. It need to be pointed out that the dual problem mentioned above only refer to the inner product operation of training samples in the optimal function seeking, for the nonlinear image data discussed in our scheme, it only needs replace the inner product operation mentioned above by kernel function K(xi, xj) to implement the nonlinear function regression.

I ff R x; y; f G x; y; f B x; yg 0 6 x < M; 0 6 y < N
denote a host digital image (color image), and fR(x, y), fG(x, y), fB(x, y) are the color component values at position (x, y). W = {w(i, j), 0 6 i < P, 0 6 j < Q} is a binary image to be embedded within the host image, and w(i, j) 2 {0, 1} is the pixel value at (i, j).

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

2087

Origin Color Image

Color Image Normalization

Significant Region Extraction

Selecting the Green Component of the Significant Region Computing the HVS Masking

NSCT

Watermarked Color Watermarked Color Image Image

Inverse Color Image Normalization

Inverse NSCT

Embedding Watermark into the Low Frequency NSCT Coefficients

Key

Watermark
Fig. 7. Watermark embedding scheme.

The digital watermark embedding scheme can be summarized as follows. Step1: Watermark preprocessingIn order to dispel the pixel space relationship of the binary watermark image, and improve the robustness of the whole digital watermark system, watermark scrambling algorithm is used at rst. In our watermark embedding scheme, the binary watermark image is scrambled from W to W1 by using Arnold transform, where

R (vi) If C R 1 xC ; yC C 0 xI ; yI , then go to (v); else, let R CR x ; y C x ; y , go to (iii). I C I C 0 1 (v) C R 0 xI ; yI is the invariant centroid of normalized red component. (vi) Select the invariant centroid as reference center, and extract rectangle area with size of S1 S2 as the signicant region for the normalized red component, as shown in Fig. 8(a).

W 1 fw1 i; j; 0 6 i < P; 0 6 j < Q g

Then, it is transformed into a one-dimensional sequence of ones and zeros as follows:

W 2 fw2 k w1 i; j; 0 6 i < P; 0 6 j < Q ; k i Q j; w2 k 2 f0; 1gg 6

Step 2: Image normalization on host color imageImage normalization technique has been used widely in computer vision and pattern recognition for a long time (Wood, 1996). The normalized image obtained from a series of geometric transformation is invariant to any afne distortions of the image. In order to improve the robustness against geometrical distortions, we apply the color image normalization on origin host image to produce the standard size normalized image. As shown in Fig. 8, the color image normalization consists of four steps. (i) Extract the red, green, and blue components from the origin color image; (ii) Perform the gray image normalization (Wood, 1996) on red component; (iii) By using the same invariant centroid of red component, perform the gray image normalization on green and blue components respectively; (iv) Combine the normalized red, green, and blue components. Step 3: The signicant region extractionAccording to the color image normalization, the normalized color image generally has the redundant characteristic, means that there is black edge after color image is normalized (Wood, 1996), as shown in Fig. 8(d). So, the signicant region is extracted from the normalized color image by using the invariant centroid, which is more suitable for embedding digital watermark. The extracting steps are explained as follows. (i) Perform smooth processing on the normalized red component by using Gaussian lter, which can eliminate effectively the noise. (ii) Compute the centroid C R 0 xI ; yI of the normalized red component, and set it as the initial invariant centroid. (iii) Calculate the invariant centroid C R 1 xc ; yc of the circular region according to invariant centroid denition. Here, (xc, yc), r are the center and radius of the circular region respectively.

We can also extract the signicant regions for normalized green component and normalized blue component, and obtain the signicant regions for normalized color image by combining the extracted three signicant regions, as shown in Fig. 8. Fig. 9 shows the color image, normalized color image and significant region under different attacks. Obviously, the signicant region of color image is robust against common image processing and geometrical distortions. Step 4: The NSCT of green component of signicant regionThe NSCT is a fully shift-invariant, multiscale, and multi-direction expansion that has better directional frequency localization and a fast implementation. The NSCT has proven to be very efcient in image processing. In order to improve the robustness against noise-like common image processing operations, the digital watermark will be embedded into the green component of signicant region in NSCT domain in this paper. After the NSCT has been applied on the green component of signicant region, the corresponding lowpass subband A is obtained. Step 5: Selecting the watermark embedding positionThe lowpass subband A is divided into 2 2 NSCT coefcients blocks Ai (i = 0, 1, . . . , S1/2 S2/2 1). By using secret key K1, P Q + H NSCT coefcients blocks are selected from the lowpass subband A. Here, the former P Q NSCT coefcients blocks Bk (k = 0, 1, . . . , P Q 1) are used for embedding digital watermark, and the latter H NSCT coefcients blocks Ck (k = P Q, P Q + 1, . . . , P Q + H 1) are used for SVR-based watermark extraction. Step 6: Digital watermark embeddingIn this paper, the digital watermark is embedded into the green component of signicant region by modifying the selected P Q NSCT coefcient blocks Bk.

bk x; y bk x; y D 2w2 k 1 V k x 0; 1; y 0; 1 k 0; 1; . . . ; P Q 1
0

where bk x; y is the modied NSCT coefcients in block Bk, bk(x, y) is the original NSCT coefcients in block Bk, and Vk is the average value of all NSCT coefcients in block Bk. D is watermark embedding strength. In this paper, we control adaptively

2088

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

Fig. 8. The color image normalization and signicant region extraction: (a) The red component, and its normalized version and signicant region, (b) The green component, and its normalized version and signicant region, (c) The blue component, and its normalized version and signicant region, (d) The color image, and its normalized version and signicant region.

the watermark embedding strength D according to the HVS masking (Qi, Zheng, & Zhao, 2008)

MF maxM L;p minM T ; F M E ; EDI EDE I

where MF is the HVS masking, p is the weight value, ranging from 0.4 to 0.5, ML is the luminance masking, MT is the texture masking, I is the original image (here, I refers to the green component of signicant region), EDE () is the edge detection

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

2089

Fig. 9. The normalized color image and its signicant region: (a) The origin color image, normalized color image and signicant region, (b) The noisy color image, normalized color image and signicant region, (c) The afne-transformed color image, normalized color image and signicant region.

operation, EDI() is the dilation operation, F() is the lter operation. Fig. 10 shows the HVS masking of the gray image Lena. Step 7: Obtaining the watermarked imageFirstly, the watermarked green component of signicant region can be obtained by performing the inverse NSCT transform with the modied lowpass NSCT coefcients and the original highpass NSCT coefcients. Secondly, the watermarked signicant region of the normalized color image can be obtained by combining the watermarked green component, and original red component and blue component. Finally, we can obtain the watermarked color image by performing the inverse color image normalization with the watermarked signicant region. 5. Watermark detection scheme According to the high correlation among different channels of the color image, a robust color image watermarking detection algorithm based on support vector regression (SVR) is proposed, which

neither needs the original host color image nor any other side information. Let

n R o G B I f x; y; f x; y; f x; y 0 6 x < M ; 0 6 y < N
denote the watermarked color image, and f x; y; f x; y; f x; y are the color component values at position (x, y). As shown in Fig. 11, the main steps of the watermark detecting procedure developed can be described as follows. Step 1: Image normalization on the watermarked color imageColor image normalization is applied on the watermarked image I* to produce the standard size normalized image. Step 2: The signicant region extractionThe signicant region is extracted from the normalized watermarked color image by using the invariant centroid.
R G B

2090

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

blocks are selected from the lowpass subbands RA* and GA*, respectively. Here, the H NSCT coefcients blocks RC k and GC k k P Q ; P Q 1; . . . ; P Q H 1 from RA* and GA* are used for training the SVR model, and P Q NSCT coefcients blocks RB k and GBk k 0; 1; . . . ; P Q 1 from RA* and GA* are used for extracting digital watermark. Step 5: SVR trainingLet the average value of all NSCT coefcients in block RC k

RV k

k P Q ; P Q 1; . . . ; P Q H 1

from lowpass subband RA* be the feature vectors for training. And let the average value of all NSCT coefcients in block GC k

GV k

k P Q ; P Q 1; . . . ; P Q H 1

from lowpass subband GA* be the training objective. Then, we can obtain the training samples as following
X k RV k ; GV k

Fig. 10. The HVS masking for gray image Lena.

k P Q ; P Q 1; . . . ; P Q H 1

Step 3: The NSCT of red component and green component of significant regionAfter the NSCT has been applied on red component and green component of signicant region, the corresponding lowpass subbands RA* and GA* are obtained. Step 4: Selecting the embedded positionThe lowpass subbands RA* and GA* are divided into 2 2 NSCT coefcients blocks RA i and GA i i 0; 1; . . . ; S1 =2 S2 =2 1. By using the same secret key K1, P Q + H NSCT coefcients

So, the SVR model can be obtained by training.Step 6: Digital watermark extractingFirstly, let the average value of all NSCT coefcients in block RB k

RV k

k 0; 1; . . . ; P Q 1

from lowpass subband RA* be the input vector. Secondly, the actual output vector can be obtained by using the well trained SVR model

GV k

k 0; 1; . . . ; P Q 1

Selecting the Red Component The Watermarked Color Image Color Image Normalization Significant Region Extraction Selecting the Green Component

NSCT and Selecting Low Frequency Constructing the SVR Training Mode NSCT and Selecting Low Frequency Digital Watermark Extraction

Watermark
Fig. 11. The watermark detecting scheme.

Fig. 12. The test images and digital watermark: (a) The test image Lena, (b) The test image Mandrill, (c) The test image Barbara, (d) The digital watermark.

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

2091

Fig. 13. The watermarked images obtained by the proposed scheme: (a) Lena (PSNR = 40.57 dB), (b) Mandrill (PSNR = 41.67 dB), (c) Barbara (PSNR = 40.71 dB).

Fig. 14. The absolute difference between origin image and watermarked image for our algorithm: (a) The absolute difference between origin image and watermarked image (Lena), (b) The absolute difference between origin image and watermarked image (Mandrill), (c) The absolute difference between origin image and watermarked image (Barbara).

Finally, the digital watermark w 2 k can be extracted by compar ing the actual output vector GV k with the GV k from lowpass sub band GA* (GV k is the average value of all NSCT coefcients in block GB k , and the rule of extracting digital watermark can be described

6. Simulation results We test the proposed watermarking scheme on the popular test images 512 512 8 bit Lena, Mandrill, and Barbara, and a 32 32 binary image is used as the digital watermark, as shown in Fig. 12. The size of the signicant region is S1 = S2 = 256, the number of training sample is 2048, and the radius-based function (RBF) is selected as the SVR kernel function, and other parameters are set respectively to c = 0.125, C = 78. Besides, the PSNR (Peak Signal-to-Noise Ratio) is used to measure the visual quality of the watermarked images. Finally, experimental results are compared with schemes in Tsai and Sun (2007) and Fu and Shen (2008).

( w 2 k

1; if GV k > GV k 0; else

k 0; 1; . . . ; P Q 1

Step 7: Watermark postprocessingAll the detected watermark bits w 2 k are rearranged to form the binary watermark image W 1 , and the watermark image

W fw i; j; 0 6 i < P; 0 6 j < Q g
can be obtained by descrambling.
Table 1 The performance test for different color image watermarking (dB). Test image Lena Mandrill Barbara Our method 40.57 41.67 40.71 Scheme (Tsai & Sun, 2007) 39.43 41.76 42.49 Scheme (Fu & Shen, 2008) 38.10 38.90 39.72

6.1. Performance test As shown in Fig. 13(a)(c) are the watermarked images (Lena, Mandrill, and Barbara) obtained by using the proposed scheme. Fig. 14 gives the absolute difference between origin image and watermarked image for our algorithm. Table 1 shows the transparency comparison results of three kinds of color image watermarking scheme.

2092

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

Fig. 15. The attacked (common image processing operations) watermarked image and extracted watermark: (a) Median ltering (3*3), BER = 0.0303, (b) Gaussian lter (3*3), BER = 0.0313, (c) Gaussian noise (0.006), BER = 0.0273, (d) Salt and Peppers noise (0.003), BER = 0.0234, (e) Sharpening, BER = 0.0225, (f) JPEG 70, BER = 0.0321, (g) JPEG 50, BER = 0.0334, (h) JPEG 30, BER = 0.0400, (i) Gaussian noise (0.006) + Sharpening, BER = 0.0449, (j) Gaussian ltering (3 3) + JPEG 70, BER = 0.0381.

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

2093

Fig. 16. The attacked (geometric distortions) watermarked image and extracted watermark: (a) Rotation (5), BER = 0.0303, (b) Rotation (45), BER = 0.0342, (c) Scaling (0.5), BER = 0.0332, (d) Scaling (2), BER = 0.0273, (e) Translation (H 30, V 0), BER = 0.0176, (f) Translation (H 0, V 30), BER = 0.0283, (g) Translation (H 5, V 15), BER = 0.0977, (h) Translation (H 20, V 20), BER = 0.1240, (i) Cropping (50%), BER = 0.1250, (j) Cropping (20%), BER = 0.1025, (k) Afne transformation[5; 1.0, 1.0; 0.3, 0.1], BER = 0.0264, (l) Afne transformation[10; 1.0, 1.0; 0.5, 0.2], BER = 0.0225, (m) Flip vertically, BER = 0.1826, (n) Flip horizontally, BER = 0.1016, (o) Rotation 5 + Scaling (2), BER = 0.0332, (p) Rotation 45 + Scaling (0.5), BER = 0.0430, (q) Translation (H 0, V 15)+ Scaling (0.5), BER = 0.0820, (r) Translation (H 0, V 15)+ Scaling (2), BER = 0.0889, (s) Rotation (5) + Translation (H 5, V 15), BER = 0.0498, (t) Scaling (0.5) + Rotation (5) + Translation (H 5, V 15), BER = 0.0635, (u) Scaling (2) + Rotation (45) + Translation (H 20, V 20), BER = 0.0709, (v) Lengthwidth ratio change (0.8, 1.0), BER = 0.0273, (w) Lengthwidth ratio change (1.2, 1.0), BER = 0.0244, (x) Rotation distortion (15), BER = 0.3613.

6.2. Robustness to various attacks Simulation results, which are obtained by the proposed watermarking scheme, for common image processing operations and geo-

metric distortions are shown in Figs. 15 and 16 respectively. Tables 2 and 3 show the results of comparison with scheme (Fu & Shen, 2008; Tsai & Sun, 2007). In this study, reliability was measured as the bit error rate (BER) of extracted watermark, its denition is

2094

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

Fig. 16 (continued)

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

2095

Fig. 16 (continued)

BER

B 100% PQ

7. Conclusion In this paper, we have described a blind color image watermarking algorithm by using the support vector regression (SVR) and nonsubsampled contourlet transform (NSCT). The extensive experimental works have shown that the proposed color image watermarking has conquered those challenging

where B is the number of erroneously detected bits, P Q is the watermark image dimensions. It is clear that the proposed scheme outperforms scheme (Fu & Shen, 2008; Tsai & Sun, 2007) under most attacks in terms of BER.

2096

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098

Fig. 16 (continued)

Table 2 The watermark detection results for common image processing operations (BER). Attacks Lena Mandrill Barbara

Proposed Scheme in Tsai Scheme in Fu Proposed Scheme in Tsai Scheme in Fu Proposed Scheme in Tsai Scheme in Fu scheme and Sun (2007) and Shen (2008) scheme and Sun (2007) and Shen (2008) scheme and Sun (2007) and Shen (2008) Median ltering (3 3) Gaussian ltering (3 3) Random noise (10) Gaussian noise (0.006) Salt and Peppers noise (0.003) Sharpening JPEG 70 50 30 Median ltering (3 3) + Gaussian noise (0.006) Gaussian noise (0.006) + Sharpening Gaussian ltering (3 3) + JPEG 70 JPEG 70 + Median lter (3 3) 0.0303 0.0313 0.0293 0.0273 0.0234 0.0225 0.0321 0.0334 0.0400 0.0244 0.0703 0.0352 0.0596 0.0557 0.0293 0.0195 0.2217 0.2539 0.3721 0.2764 0.0496 0.0381 0.0322 0.0264 0.0273 0.0244 0.1533 0.2080 0.2471 0.0430 0.0049 0.0107 0.0059 0.0234 0.0205 0.0449 0.0203 0.0293 0.0322 0.0137 0.2422 0.1865 0.0947 0.1338 0.0850 0.0596 0.2002 0.2705 0.3652 0.3037 0.1162 0.0967 0.1104 0.1035 0.1221 0.1309 0.2158 0.2969 0.3320 0.1279 0.0234 0.0225 0.0195 0.0215 0.0164 0.0273 0.0254 0.0244 0.0283 0.0186 0.1465 0.0928 0.1035 0.1416 0.0518 0.0635 0.2559 0.4258 0.4307 0.2813 0.0859 0.0605 0.0547 0.0898 0.0684 0.0928 0.2109 0.2549 0.2705 0.1396

0.0449 0.0381 0.0264

0.1621 0.3213 0.3281

0.0420 0.1904 0.1855

0.0811 0.0234 0.0195

0.1445 0.3506 0.3779

0.1348 0.2324 0.2686

0.0547 0.0303 0.0196

0.2139 0.3027 0.3174

0.1270 0.2217 0.2422

geometric distortions, such as rotation, translation, scaling, and lengthwidth ratio change etc. Also, the digital watermark can resist some common image processing operations.

Drawbacks of the proposed color image watermarking scheme are related to the computation time for SVR training, and performing NSCT and image normalization. Future work will focus on eliminating these drawbacks.

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098 Table 3 The watermark detection results for geometric attacks (BER). Attacks Lena Mandrill Barbara

2097

Proposed Scheme in Tsai Scheme in Fu Proposed Scheme in Tsai Scheme in Fu Proposed Scheme in Tsai Scheme in Fu scheme and Sun (2007) and Shen (2008) scheme and Sun (2007) and Shen (2008) scheme and Sun (2007) and Shen (2008) Rotation 5 45 Scaling 2 Translation (H 5, V 0) (H 30, V 0) (H 0, V 5) (H 0, V 30) (H 20, V 20) (H 5, V 15) Cropping (x 10%, y 10%) (x 20%, y 20%) Lengthwidth (0.9, 1.1) ratio change (1.2, 1.0) Flip vertically Flip horizontally Afne transformation [10; 1.0, 1.0; 0.5, 0.2] Afne transformation [10; 1.0, 1.0; 0.5, 0.2] Scaling 2 + Translation (H 5, V 0) Rotation 5 + Scaling 2 Rotation 5 + Translation (H 5, V 15) Rotation 45 + Scaling 2 + Translation (H 20, V 20) 0.0303 0.0342 0.0273 0.0332 0.0176 0.0186 0.0283 0.1240 0.0977 0.0781 0.1025 0.0420 0.0244 0.1826 0.1016 0.0225 0.0273 0.0596 0.0332 0.0498 0.0709 0.5273 0.5000 0.5059 0.4971 0.4791 0.5049 0.5049 0.4922 0.5088 0.0205 0.0313 0.5098 0.5049 0.4883 0.4922 0.4785 0.5059 0.5215 0.4609 0.4951 0.5117 0.5244 0.5264 0.5371 0.4980 0.4922 0.5059 0.5059 0.5146 0.5068 0.0264 0.0361 0.5352 0.5176 0.5137 0.5449 0.5225 0.5215 0.5283 0.4371 0.5234 0.5322 0.0127 0.0164 0.0137 0.0244 0.0166 0.0145 0.0145 0.0605 0.0479 0.0684 0.0977 0.0186 0.0166 0.0684 0.0801 0.0137 0.0234 0.0273 0.0234 0.0479 0.1318 0.5078 0.4971 0.4951 0.5010 0.5254 0.4863 0.4863 0.5176 0.5039 0.1113 0.1162 0.5215 0.4834 0.4941 0.4893 0.4951 0.5215 0.4902 0.5107 0.5000 0.5283 0.5391 0.5020 0.4464 0.4922 0.5176 0.5205 0.5205 0.4980 0.5020 0.1240 0.1309 0.5303 0.5205 0.4980 0.4883 0.5166 0.5000 0.5205 0.5361 0.5186 0.5352 0.0254 0.0244 0.0303 0.0215 0.0205 0.0195 0.0195 0.1201 0.1201 0.1230 0.1367 0.0234 0.0244 0.1270 0.1074 0.0195 0.0254 0.0713 0.0254 0.1240 0.1221 0.4863 0.4854 0.5283 0.5000 0.4824 0.4961 0.4961 0.4902 0.4805 0.0684 0.0603 0.4941 0.5127 0.5049 0.4863 0.4893 0.4922 0.5176 0.4980 0.5068 0.4902 0.5283 0.5449 0.5215 0.5303 0.5195 0.5137 0.5137 0.5186 0.5117 0.1602 0.1670 0.5186 0.5327 0.4941 0.4941 0.4912 0.5010 0.5195 0.5273 0.5049 0.4761

References
Arash, A., & Shohreh, K. (2008). Color PCA eigenimages and their application to compression and watermarking. Image and Vision Computing, 26, 878890. Behnia, S., Teshnehlab, M., & Ayubi, P. (2010). Multiple-watermarking scheme based on improved chaotic maps. Communications in Nonlinear Science and Numerical Simulation, 15(9), 24692478. Boato, G., & Conotter, V. (2009). Watermarking robustness evaluation based on perceptual quality via genetic algorithms. IEEE Transactions on Information Forensics and Security, 4(2), 207216. Cunha, A. L., Zhou, J., & Do, M. N. (2006). The nonsubsampled contourlet transform: Theory, design, and applications. IEEE Transactions on Image Processing, 15(10), 30893101. Deng, C., Gao, X., Li, X., & Tao, D. (2009). A local tchebichef moments-based robust image watermarking. Signal Processing, 89(8), 15311539. Dong, P., Jovan, G. B., Yongyi, P. G., Yang, N., & Franck, D. (2005). Digital watermarking robust to geometric distortions. IEEE Transactions on Image Processing, 14(12), 2152140. Fu, Y. G., & Shen, R. M. (2008). Color image watermarking scheme based on linear discriminant analysis. Computer Standard & Interfaces, 30(3), 115120. Gao, X. B., Deng, C., & Li, X. L. (2010). Geometric distortion insensitive image watermarking in afne covariant regions. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, 40(3), 278286. Guo, J., Liu, Z. J., & Liu, S. T. (2007). Watermarking based on discrete fractional random transform. Optics Communications, 272(2), 344348. Lee, M. H., & Kwon O. J. (2008). Color image watermarking based on DS-CDMA using hadamard kernel. In ICACT 2008, February 1720 (pp. 15921597). Lee, H.-Y., Lee, C.-H., & Lee, H.-K. (2007). Geometrically invariant watermarking: Synchronization through circular Hough transform. Multimedia Tools and Applications, 34(3), 337353. Lian, S., Kanellopoulos, D., & Ruffo, G. (2009). Recent advances in multimedia information system security. Informatica, 33(1), 324. Li, C.-H., Ling, H. -F., & Lu, Z. -D. (2007). Semi-fragile watermarking based on SVM for image authentication. In IEEE international conference on multimedia and expo, 2007, 25 July 2007 (pp. 12551258). Li, L.-D., & Guo, B.-L. (2009). Localized image watermarking in spatial domain resistant to geometric attacks. International Journal of Electronics and Communications, 63(2), 123131. Li, L. D., Guo, B. L., & Pan, J. (2010). Robust image watermarking using scale-space feature points in contourlet domain. IEEE Transactions on Information Forensics and Security. Lin, S. D., Shie, S. C., & Guo, J. Y. (2010). Improving the robustness of DCT-based image watermarking against JPEG compression. Computer Standards & Interfaces, 32(12), 5460.

Liu, K.-C. (2010). Wavelet-based watermarking for color images through visual masking. AEU International Journal of Electronics and Communications, 64(2), 112124. Liu, Y., Zheng, D., & Zhao, J. (2007). An image rectication scheme and its applications in RST invariant digital image watermarking. Multimedia Tools and Applications, 34(1), 5784. Peng, Z. N., & Liu, W. B. (2008). Color image authentication based on spatiotemporal chaos and SVD. Chaos, Solitons & Fractals, 36, 946952. Pham, V.-Q., Miyaki, T., Yamaski, T., & Aizama, K. (2008). Robust object-based watermarking using feature matching. IEICE Transactions on Information and Systems, E91-D(7), 20272034. Qi, H. Y., Zheng, D., & Zhao, J. Y. (2008). Human visual system based adaptive digital image watermarking. Signal Processing, 88, 174188. Singhal, N., Lee, Y. Y., & Kim, C. S. (2009). Robust image watermarking using local zernike moments. Journal of Visual Communication and Image Representation, 20(6), 408419. Tai, W. L., Yeh, C. M., & Chang, C. C. (2009). Reversible data hiding based on histogram modication of pixel differences. IEEE Transactions on Circuits and Systems for Video Technology, 19(6), 906910. Tsai, H. H., & Sun, D. W. (2007). Color image watermark extraction based on support vector machines. Information Sciences, 177(2), 550569. Tsui, T. K., Zhang, X.-P., & Androutsos, D. (2008). Color image watermarking using multidimensional Fourier transforms. IEEE Transactions on Information Forensics and Security, 3(1), 1628. Vapnik, V. (1995). The nature of statistical learning theory. New York: SpringerVerlag. Wang, X., Wu, J., & Niu, P. (2007). A new digital image watermarking algorithm resilient to desynchronization attacks. IEEE Transactions on Information Forensics and Security, 2(4), 655663. Wang, X.-Y., Xu, Z.-H., & Yang, H.-Y. (2009). A robust image watermarking algorithm using SVR detection. Expert Systems with Applications, 36(5), 90569064. Wood, J. (1996). Invariant pattern recognition: A review. Pattern Recognition, 29(1), 117. Wu, J.-Z., & Xie, J.-Y. (2006). An adaptive watermarking scheme utilizing support vector machine for synchronization. Journal of Shanghai Jiaotong University, 40(3), 481484. Xiang, S., Kim, H. J., & Huang, J. (2008). Invariant image watermarking based on statistical features in the low-frequency domain. IEEETransactions on Circuits and Systems for Video Technology, 18(6), 777790. Xing, Y., & Tan, J. Q. (2007). A color watermarking scheme based on block-SVD and arnold transformation. In Proceedings of workshop on digital media and its application in museum & Heritage, Chongqing, China, 2007 (pp. 38). Xing, Y., & Tan, J. (2010). A color image watermarking scheme resistant against geometrical attacks. Radio Engineering, 19(1), 6267. Xin, Y., Liao, S., & Pawlak, M. (2007). Circularly orthogonal moments for geometrically robust image watermarking. Pattern Recognition, 40(12), 37403752.

2098

P.-P. Niu et al. / Expert Systems with Applications 38 (2011) 20812098 Zheng, J. B., & Feng, S. (2009). A color image watermarking scheme in the associated domain of DWT and DCT domains based on multi-channel watermarking framework. In IWDW 2008. LNCS (Vol. 5450, pp. 419432). Springer. Zheng, D., Liu, Y., Zhao, J., & El Saddik, A. (2007). A survey of RST invariant image watermarking algorithms. ACM Computing Surveys, 39(2), 191. Zheng, D., Wang, S., & Zhao, J. (2009). RST invariant image watermarking algorithm with mathematical modeling and analysis of the watermarking processes. IEEE Transactions on Image Processing, 18(5), 10551068.

Zhang, C., Cheng, L. L., Qiu, Z. D., & Cheng, L. M. (2008). Multipurpose watermarking based on multiscale curvelet transform. IEEE Transactions on Information Forensics and Security, 3(4), 611619. Zhang, X. P., Li, K., & Wang, X. F. (2008). A novel look-up table design method for data hiding with reduced distortion. IEEE Transactions on Circuits and Systems for Video Technology, 18(6), 769776. Zhang, Li., Qian, G.-B., Xiao, W.-W., & Ji, Z. (2007). Geometric invariant blind image watermarking by invariant Tchebichef, moments. Optics Express, 15(5), 22512261.