This action might not be possible to undo. Are you sure you want to continue?

Contents lists available at ScienceDirect

**Expert Systems with Applications
**

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

**A blind watermarking method using maximum wavelet coefﬁcient quantization
**

Wei-Hung Lin a, Yuh-Rau Wang b,*, Shi-Jinn Horng a,c, Tzong-Wann Kao d, Yi Pan e

a

Department of Computer Science and Information Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan Department of Computer Science and Information Engineering, St. John’s University, Taipei, Taiwan Department of Electronic Engineering, National United University, Miao-Li, Taiwan d Department of Electronic Engineering, Technology and Science Institute of Northern Taiwan, Taipei, Taiwan e Department of Computer Science, Georgia State University, Atlanta, GA 30302-4110, USA

b c

a r t i c l e

i n f o

a b s t r a c t

This paper proposes a blind watermarking algorithm based on maximum wavelet coefﬁcient quantization for copyright protection. The wavelet coefﬁcients are grouped into different block size and blocks are randomly selected from different subbands. We add different energies to the maximum wavelet coefﬁcient under the constraint that the maximum wavelet coefﬁcient is always maximum in a block. The watermark is embedded the local maximum coefﬁcient which can effectively resist attacks. Also, using the block-based watermarking, we can extract the watermark without using the original image or watermark. Experimental results show that the proposed method is quite robust under either non-geometry or geometry attacks. Ó 2009 Elsevier Ltd. All rights reserved.

Keywords: Watermark Blind Copyright protection Maximum coefﬁcient quantization Wavelet

1. Introduction The rapid growth of the Internet and digital media manifests itself in widespread public forms such as the digital image, the MPEG, and so on, because digital media are easy to copy and transmit. Many researchers are aware of the issues of copyright protection, image authentication, proof of ownership, etc. Hence, there are many solutions that have been proposed. The watermarking technique is one of the solutions. This technique embeds information so that it is not easily perceptible; that is, the viewer cannot see any information embedded in the contents. A watermarking technique is referred to as blind if the original image and watermark are not needed during extraction (Chen, Horng, & Lee, 2005; Dugad, Ratakonda, & Ahuja, 1998; Ganic & Eskicioglu, 2004; Inoue, Miyazaki, & Yamamoto, 1998; Tsai, Yu, & Chen, 2000). There are several important issues in the watermarking system. First, the embedded watermark should not degrade the quality of the image and should be perceptually invisible to maintain its protective secrecy. Second, the watermark must be robust enough to resist common image processing attacks and not be easily removable; only the owner of the image ought to be able to extract the watermark. Third, the blindness is necessary if it is difﬁcult for us to obtain the original image and watermark.

* Corresponding author. E-mail addresses: D9515013@mail.ntust.edu.tw (W.-H. Lin), yrwang@mail.sju.edu.tw (Y.-R. Wang), horngsj@yahoo.com.tw (S.-J. Horng), tkao@tsint.edu.tw (T.-W. Kao), pan@cs.gsu.edu (Y. Pan). 0957-4174/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.03.060

Spread spectrum communication is robust against many types of interference and jamming (Pickholtz, Schilling, & Milstein, 1982). In the following paragraphs, we will brieﬂy review some proposed watermarking approaches which are based on the frequency domain (discrete cosine transform (DCT) and discrete wavelet transform (DWT)). Cox, Kilian, and Leighton (1996, 1997) suggested inserting the watermark into the perceptually signiﬁcant portion of the whole DCT-transformed image, wherein a predetermined range of low frequency components excludes the DC component. This watermarking scheme has been shown to be robust against common attacks, such as compression, ﬁltering, cropping, etc. Kwon, Kim, and Park (1999) embedded the watermark in the variable DCT block size. The block size is determined according to the characteristics of the region in the spatial domain. Langelaar and Lagendijk (2001) proposed a blind watermarking approach called differential energy watermarking. A set of several 8 Â 8 DCT blocks are composed and divided into two parts to embed a watermark bit. The high frequency DCT coefﬁcients in the JPEG/ MPEG stream are selectively discarded to produce energy differences in the two parts of the same set. Although different watermarking approaches are proposed, the DWT approach remains one of the most effective and easy to implement techniques for image watermarking (Meerwald & Uhl, 2001). The most important issue in DWT-based image watermarking is how to choose the coefﬁcients to be embedded. In Dugad et al. (1998), Kim and Moon (1999), Kwon, Ban, and Ha (2001), Wang, Chen, and Cheng (2000), Wang and Huo (1997), and Wang, Su, and Kuo (1998), the watermark is embedded the signiﬁcant coefﬁcient which is selected from the wavelet coefﬁcients. Wang

and each block is used to embed one watermark bit. The block-based signiﬁcant coefﬁcient As mentioned in Section 1. We embed a watermark in different subbands and every block will be used to embed either the watermark bit 0 or bit 1.. Wang & Lin. HLLÀ1. LHL. adding the Human Visual System (HVS) to effectively resist geometric attack. Each subblock is separated into two areas. since the subbands can easily be eliminated. namely up and low. 2005. Some researches embed a watermark using block-based DWT. In Shapiro (1993). The scan order of wavelet coefﬁcients and the block-based signiﬁcant coefﬁcient 2. One of the four trees is quantized according to the binary value of the two embedded watermark bits. LHLÀ1. and the resulting value is still the maximum in the block. 1997). Experimental results show that the proposed method decreases the distortion of the host image and is effectively robust against JPEG compression. one of the two areas in each subblock is quantized. according to a sequence of information bits. The scan order of wavelet coefﬁcients A host image of size IW by IH is transformed into wavelet coefﬁcients using the L-level discrete wavelet transform (DWT). and HH1 as shown in Fig. For this reason. 1. Temi. watermarking is more robust when the watermark bits are embedded in the signiﬁcant coefﬁcients of the whole frequency domain. then the block has been embedded with a watermark bit 1. Tsai & Lin. According to a watermark bit. Fig. In the extraction process.. 2. according to multithreshold wavelet coding (MTWC) (Wang & Kuo. 2007. Liang. One of the two trees is quantized with respect to a quantization index. these methods cannot efﬁciently resist Fig. The watermark is embedded in the high subband of each block which is transformed into the DWT according to its classiﬁcation. HHL. & Lasakul. HH1 subbands is also not suitable. Hsieh. we brieﬂy introduce the scan order of wavelet coefﬁcients and the block base signiﬁcant coefﬁcient. Wu and Huang (2007) improved Wang and Lin’s method using minimum mean to extract the watermark. The problem is that the order of extracting signiﬁcant coefﬁcients in the extraction process should be exactly the same as those in the embedding process. Choomchuay.2. we use the local maximum wavelet coefﬁcient quantization.1. especially for the attacked watermarked image. 1998. A super-tree which consists of four wavelet trees is divided into ﬁve subblocks.. in this paper. the wavelet coefﬁcients of a host image are grouped into blocks of variable size. and Fig. and both trees exhibit a large statistical difference between the quantized tree and the unquantized tree. the conclusions are summed up in Section 5. with each of size n Â n. The watermark is added by quantizing the signiﬁcant coefﬁcient in the signiﬁcant band using different weights. the successive subband quantization (SSQ) was adopted to search for the signiﬁcant coefﬁcients. Lin et al. Wang et al. The watermark is embedded in the wavelet coefﬁcients in the middle and lower subbands of a block in each image.. The other considers rectangular blocks of coefﬁcients. we embed the watermark bits into the LHL and HLL subbands. The watermark is embedded by using the mean and the variance of a subband to modify the wavelet coefﬁcient of a block. Tsai and Lin (2007) proposed a structure based wavelet tree quantization. . 3 illustrates the extraction procedure. we have 3 Â L + 1 subbands. HHLÀ1. But it is not easy to use blind detection to ﬁnd the permutations in the original order of the signiﬁcant coefﬁcients. / Expert Systems with Applications 36 (2009) 11509–11516 et al. Tseng. . 2004. According to the characteristics of this scan order. Wu & Huang. With an L-level decomposition. By adjusting the local maximum wavelet coefﬁcient and maintaining the maximum wavelet coefﬁcient always maximum to present a binary watermark. Finally. But these methods (Li et al. 2001. . 2006. . and Wen (2004) divided the original image into n Â n blocks and transformed these into a DWT domain. Kheliﬁ.11510 W. The signiﬁcant subbands must be transmitted in the order of LLL. 1. . they were unsuitable for the blind watermarking. Transmission of wavelet subbands by scan order. Although there are many block-based watermarking methods are proposed. the block has been embedded the watermark bit 0.. The original image is separated into m blocks. Davoine (2000) proposed two watermarking methods. then every block is decomposed into a wavelet domain.. and Kurugollu (2005) proposed an adaptive blind watermarking technique based on the DWT. 2. The proposed method is described in Section 3.. Embedding a watermark in the HHL. otherwise. the difference can later be used for watermark extraction. 1997) use signiﬁcant coefﬁcients which are selected from global coefﬁcients are showed that can effectively resist attacks. Wang and Lin (2004) proposed a wavelet-based blind watermarking scheme and each watermark bit is embedded using two trees. 2000. and Niu (2006) improved Wang and Lin’s (2004) method. One is the watermark embedded by modifying the triplets of signiﬁcant coefﬁcients. Hsieh et al. HL1. 2006. The experimental results and experimental analysis are given in Section 4. HLL. low-pass ﬁltering and Gaussian noise. we embed a watermark bit 1 by increasing the energy of the local maximum wavelet coefﬁcient in a block and embed a watermark bit 0 by decreasing the energy of the local maximum wavelet coefﬁcients in a block. LH1. The original image is separated into non-overlapping blocks classiﬁed as uniform or non-uniform blocks using a JND-based classiﬁer. Li. 2007) cannot effectively resist low-pass ﬁltering such as median ﬁltering or Gaussian ﬁltering. Lien & Lin. 1998. HHLÀ1. Wang & Huo. Huang and Yang (2004) proposed a watermarking algorithm based on the DWT. the signiﬁcant wavelet coefﬁcients are transmitted by the signiﬁcant frequency subbands. Wang. for example by lossy compression. and the PSNR value of a watermarked image is greater than 40 dB. if we subtract the energy from the local maximum coefﬁcient in a block. and Huang (2001) proposed a watermarking method based on the qualiﬁed signiﬁcant wavelet tree (QSWT). (1998) proposed a watermarking method. So. Bouridane. This paper is organized as follows: In Section 2.-H. Previous researches (Dugad et al.. The LLL frequency band is found to be unsuitable to be modiﬁed since it is a low frequency band which contains important information about an image and easily causes image distortions. 2 illustrates the watermark embedding procedure using 3-level DWT. Lien and Lin (2006) improved Wang and Lin’s (2004) method by using four trees to represent two watermark bits in order to improve visual quality. Zhang.

. either Block0. A pseudorandom . for example. AvailCoeff = IL À NB. 3. Fig. J = 1.1. the selected blocks form a set of blocks called BlockQue. we will use these blocks to embed the watermark. respectively. The watermark embedding procedure: embed a watermark bit by quantizing the maximum wavelet coefﬁcient in a block.j.j wavelet coefﬁcients as the HLL subband. or Block1. each of BlocksN blocks (see Algorithm 1).W. Nw denotes the number of watermark bits which are embedded. To avoid BlocksN < Nw and to assure that there are enough blocks for embedding watermark bits. The proposed method 3. 2.j is selected.j.j is used. BlocksN = BlocksN + 1. Algorithm 1.j is chosen. When embedding a watermark bit. / Expert Systems with Applications 36 (2009) 11509–11516 11511 Fig. as does the HLL subband. the LHL subband totally has IL wavelet coefﬁcients. we have BlocksN blocks in LHL and HLL subbands.3. We deﬁne È É E ¼ ej jej 2 f0. NB. 1g.j . For example. where ð2Þ È É BlockQue ¼ BlockQuej jBlockr.j denotes the block size of these blocks. 1 6 j 6 BlocksN . then Block0. otherwise Block1. j++. .-H. In the jth block. Lin et al. ð1Þ where r = 0 denotes the LHL subband and r = 1 denotes the HLL subband. NB. r ¼ ej . since a watermark bit is embedded in the coefﬁcient of a block of size n Â n and the signiﬁcant coefﬁcients are not considered. In the next section. . if we set NB = 7. IW and IH are the width and height of the host image. . Watermark extraction procedure. 3. We shufﬂe the BlockQue blocks in a pseudorandom manner.j 5 NB = 7. 1 6 j 6 BlocksN : ð3Þ If ej = 0.j or Block1.j 2 {2. each time we must select a block from either Block0. Hence.j > = 0) { NB. To solve the above problems. 1g and 1 5 j 5 BlocksN . Hence. initially. then we quantize the maximum coefﬁcient of the block to embed a watermark bit. } where IL is the number of wavelet coefﬁcients in the L-level subband.(IL/Nw)}. In order to decrease the probability that the watermark may be tampered with. While (IL À NB.j to construct a block. the LHL subband has NB. attacks.j is an integer which ranges from 2 to NB. j k IL ¼ IW =2L Â IH =2L : AvailCoeff = IL. we deﬁne Blockr. The preprocess In the previous section.j to embed a watermark bit.j ¼ NB.j for r f0. where NB denotes the maximum number of wavelet coefﬁcients in a block. Both LHL and HLL subbands use the same NB. BlocksN = 0. we group the variable number of the wavelet coefﬁcients from the LHL and HLL subbands into a block. 2 5 NB 5 (IL/Nw). then 2 5 NB. NB must be less than or equal to IL/Nw.

although the value of maxi Â T1 is still less than the value of seci Â T2. Eq. i Otherwise. & Chang. The other condition is thatai is set to Meanembed Â Tq. r2 ¼ av g 0i =Meanblock . which may cause error in the watermark extraction process. otherwise. in order to subjectively verify the ownership of an image by extracting a watermark. otherwise. and we use 3-level DWT. the average maximum wavelet coefﬁcient of the watermarked image is obtained. Now. r1 ¼ max0i =Meanextract . According to the watermark bits embedded later. we invoke a pseudorandom function to shufﬂe Nw bits by using different seeds. Fig. Let Tq be a scale parameter. the value of ai is not added to Meanembed when the block has an embedded watermark bit 1. For example. if the watermark bit is 1. When the value of ai is subtracted from the local maximum wavelet coefﬁcient max0i of a watermarked image. Since the value of avgi is unchanged. We represent each watermark bit as either 1 or 0. we seek the local maximum wavelet coefﬁcient maxi and the local second maximum wavelet coefﬁcient seci. if we set Tq = 0. 4 shows that the larger the Tq is. the latter is used for subjective detection (Potdar. a binary logo image is used. ð7Þ a0i ¼ ðjMeanblock j þ r1 þ r2 Þ. we generate BlockQue from the watermarked image the same as in the embedding process. (4). Each time we embed a watermark bit. Finally. maxi . The former is used for objective detection. maxnew ¼ maxi À ai . During the embedding process. Note that we use these two generated random numbers to denote the indices of two blocks. Finally the value of a0i is obtained by Eq. however. i Nw i¼1 ð10Þ where maxnew is the new value of maxi. T2 = 0. the image distortion increases as well. the magnitude of the local maximum wavelet coefﬁcient is decreased. / Expert Systems with Applications 36 (2009) 11509–11516 order of the numbers from Blocks1 to BlocksN can be obtained by repeating two random numbers bBlocksN/2 c times. Here.3. ð11Þ . Nw i¼1 ð4Þ & Mi ¼ if ith watermark bit is 1. If av g 0i is negative. Then. then the corresponding PSNRs of all the images are greater than 40 dB. and the maxi is quantized as where av g 0i is the average coefﬁcient value of the ith block excluding max0i . the maxi is quantized as Nw 1 X max0i : Nw i¼1 ð9Þ In Eq. Watermark embedding Two types of techniques are proposed for embedding a pre-deﬁned watermark into an image. the value of Meanblock in all Nw blocks is computed by Eq. it is very important to ﬁnd the suitable ai value. Lena and Peppers images (512 Â 512 pixels. In Fig. each time we generate two random numbers by using the same seed and taking modulo (BlocksN +1). (10). The later value will be referred to in order to quantize the local maximum wavelet coefﬁcient.11512 W. but sometimes the value of avgi is too close to zero or negligible. We also show the average PSNRs of other 12 commonly used images obtained from Petitcolas (1997). we set T1 = 0. and T2 = 0. Hence we must take both avgi and Meanembed into consideration. Clearly. In the preprocess stage of the extraction process. A binary watermark W of size Nw (5BlocksN) bits will be embedded. the block is embedded with a watermark bit 0. Hence we quantize maxi with a reference value ai which is represented as where is the new value of seci. One is pseudo random Gaussian sequence. the value is transformed using the absolute value. So the value of Meanextract is larger than Meanembed. Meanblock ¼ maxnew i ¼ maxi þ ai . maxi Â T 1 . if the value of max0i is still the local maximum in a block.2. Then the corresponding two blocks indexed by these two generated random numbers are exchanged. Therefore.3. Since we only alter the local maximum wavelet coefﬁcient. By reducing the values of maxi and seci. the more robust the watermark will be. In the experiment. We shufﬂe the blocks using the same seed as used in the embedding process. the magnitude of the local maximum wavelet coefﬁcient is increased if the watermark bit is 1. ai ¼ Maximum jav g i j. 3. Lin et al. if the two generated random numbers are 123 and 33. it is set to j avgij. To extract the watermark. i secnew i ð8Þ Meanembed ¼ where Nw 1 X Mi . One is that ai is set to avgi. maxi is the local maximum wavelet coefﬁcient of the ith block. there is a high probability that the block has been embedded a watermark bit 1. The issue is how to tradeoff between the robustness and the quality of the images. we use avgi to quantize maxi. The value of maxi Â T1 will be explained later. the remaining coefﬁcients are unchanged. the accuracy in the extraction process can be increased.1 to 1. hence the value of ai also cannot be obtained. Nw = 512. we must use another value to substitute for Meanembed Â Tq. ranging from 0. if a block is embedded with a watermark bit 1. the value of ai can be found. otherwise: Meanembed denotes the average value of the local maximum wavelet coefﬁcients in all Nw blocks. 3. Meanembed Â T q : È É ð5Þ Meanextract ¼ Originally.1. ai is obtained from either avgi or Meanembed Â Tq.-H. then block 123 and block 33 are exchanged. In the proposed method. In Eq. In Eq. in the meanwhile. and the watermark bit 1 can be extracted in the extraction process. (5).3. i where ai = maxi Â (1 À T1). in the experiment. an original image or an original watermark logo is not required for the extraction process. the reference value ai is set to Meanembed Â Tq. 8 bits/pixel) with different Tq values. The decoder design In the proposed method. (11). 1 5 i 5 Nw. the energy ai is added to the local maximum wavelet coefﬁcient. (9). we plot the peak signal-to-noise ratio (PSNR) (see Eq. 2005). In this paper. Han. (7) can be rewritten as maxnew ¼ maxi Â T 1 : i Suppose that maxnew < seci . 4.1.7. where T1 = 0. In the extraction process. some values are calculated in a process that is similar to the embedding process. ð6Þ Nw 1 X av g 0 . Both T1 and T2 are threshold values. seci is then quantized as i secnew ¼ seci Â T 2 . (5). we select Nw nonoverlapping blocks from BlockQue and compute the average local maximum coefﬁcient of all the Nw blocks using Eq. the value of Meanembed cannot be obtained. This characteristic can be used in the extraction process. the other is binary image or gray image. There are two conditions if the block is embedded with a watermark bit 1. If the value of j avgij is less than Meanembed Â Tq. The watermark bit is either 1 or 0. (14)) for the selected Baboon. We compute the average coefﬁcient value avgi of the ith block excluding the local maximum wavelet coefﬁcient maxi.

PSNR ¼ 10 Â log10 1 IH ÂIW 255 Â 255 dB. and sharpening.3. Goldhill. The above is also the case applied for r2.-H. Fig. After extracting the watermark. We predetermine the scale parameter Tq at 0. y) and g(x. y) are the values located at coordinates (x. it is set to À1. jMeanblock j þ r1 þ r2 : 0 i ð12Þ The watermark can be obtained by Eq. j) is either À1 or 1. / Expert Systems with Applications 36 (2009) 11509–11516 11513 Fig. So the value of w(i. and Peppers (512 Â 512 pixels. and the threshold values T1 and T2 as 0. histogram equalization. jÞ Â w0 ði. 8 bits/pixel). we list the PSNR formula as follows: where wh and ww are the height and width of the watermark. 4. j) is set in the same way. and Peppers) are 42.W. y) of the original image. if ðmax0i À a0i Þ P sec0i . j) 0 Â w (i. Watermark extraction Eq. respectively. the watermark image can be extracted. Nongeometric attacks include JPEG compression. Here w(i. low-pass ﬁltering. 3. and 42. PSNR. only the Peppers image is shown. j) of the original watermark and the extracted watermark. 6 shows the unattacked watermarked image and the extracted result. respectively. wh Â ww i¼0 j¼0 ð15Þ È É a ¼ Maximum av g 0i . where r1 and r2 are biased. j) and w0 (i. (13). 5 shows the original image and the binary watermark.. yÞ À g ðx.1 and 0. is greater than the value r1 which is obtained from the block embedded with a watermark bit 0 in the extraction process. 0.7.. PIH À1 PIW À1 2 x¼0 y¼0 ½f ðx. then the block is embedded with a watermark bit 1. (a) The original image of Peppers of size 512 Â 512 and (b) the original binary watermark of size 32 Â 16. namely Lena. Here we consider both geometric and nongeometric attacks. w0 (i. 2005) are used to simulate these attacks. then it will be greater than the value of av g 0i which is obtained from the block embedded with a watermark bit 0 in the extraction process. which is obtained from the block embedded with a watermark bit 1. respectively.53 dB. max0i À a0i is greater than or equal to sec0i Þ.4. They were obtained from Petitcolas (1997). jÞ. By using Eqs. Goldhill. and the attacked image.e. 1997) and PhotoImpact 11 software (Ulead Systems. the remainder value of max0i is still the local maximum in the ith block (i. It is deﬁned as Fig. We used three images for the experiments. For the sake of completeness. otherwise. respectively. otherwise: ð13Þ If after the value of a0i is subtracted from max0i . 4. Experimental results The peak signal-to-noise ratio (PSNR) is used to evaluate the quality between an attacked image and the original image. watermarkbit ¼ & 1. respectively. The value r1. 42. For brevity. The PSNRs of the three watermarked images (Lena.02. The tools of the Stirmark benchmark (Petitcolas. Lin et al. the normalized correlation coefﬁcient (NC) is computed using the original watermark and the extracted watermark to judge the existence of the watermark and to measure the correctness of an extracted watermark. j) are the watermark bits located at coordinates (i. f(x. w(i. The Tq value vs. Fig. if av g 0i is obtained from the block embedded with a watermark bit 1.58. (11)–(13) on the remaining Nw À 1 blocks. Inc. yÞ ð14Þ where IW and IH are the height and width of the image. the block is embedded with a watermark bit 0. since the value of seci is somewhere between maxi and avgi. Otherwise. . (11) is rearranged as follows: NC ¼ wh À1 ww À1 X X 1 wði. j) is set to 1 if it is watermark bit 1. L = 3. 5.

or average ﬁlter with a mask of size 5 Â 5.89 .95 (5 Â 5) 0. the positions of pixels will be shifted.92 0.53 0. Moreover. 40.76 0. and the results are shown in Table 1. the results are also shown in Table 3. 80. Their watermarking approaches are blind and their methods are based on wavelet-tree quantization.67 25 0. for example median ﬁlters. 20. 70. Table 1 Normalized correlation coefﬁcients (NC) after attacked by JPEG compression with the quality factors (QF) 10.98 1. and histogram equalization of the three test images. JPEG is one of the most frequently used formats on the Internet and in digital cameras.81 0. although an image is rotated in a small degree. our proposed method is better than theirs in robustness.90 0. 6.99 100 1. / Expert Systems with Applications 36 (2009) 11509–11516 Fig.80 (7 Â 7) 0. QF (a) Lena NC EXtracted watermark (b) Goldhill NC EXtracted watermark (c) Peppers NC EXtracted watermark 10 0.85 0. 25.00 Table 2 Normalized correlation coefﬁcients (NC) after attacked by median ﬁlter (3 Â 3. 7 Â 7). 1997).58 0. Gaussian noise and cropping are used.88 Gaussian ﬁlter Average ﬁlter (3 Â 3) 0.90 0. and then open the above image of pixels 256 Â 256 to resize them back to pixels 512 Â 512.96 0.66 0.97 0. 30. (a) Watermarked Peppers with PSNR=42. (b) Goldhill.85 40 0. sharpening. the resulting images are blurred or sharpened at the edge (Gonzalez & Woods.53 0.77 0.47 0.74 30 0.37 0.95 70 0.-H. but the quality of the resulting image is signiﬁcantly reduced. 35. 7 Â 7). the results are shown in Table 4.72 0. and 100 on (a) Lena.93 0. rotation. It is then difﬁcult to recover the original image from a rotational image.35 0.69 0. an image of 1/4 size is cropped via PhotoImpact 11 software.79 Sharpening Histogram equalization 0. 2002). scaling.00 0. The result of grouping coefﬁcients into blocks for a rotation image will be quite different from that of the original image. and Lien and Lin’s (2006) methods using the Lena image. For other nongeometric attacks (Petitcolas.46 0.11514 W. and shown in Table 2. 4. Gaussian ﬁltering.76 (7 Â 7) 0. Li et al. When the quality factor is decreased from 100. Experimental analysis We compare the proposed method to Wang and Lin’s (2004). (b) The extracted watermark with NC = 1.48 0. Though the rotation attack does not cause the image serious visual distortion. 5 Â 5.30 0.88 0. hence the size of the cropped image is still 512 Â 512.95 0. Varied quality factors (Petitcolas. 15. The noise variation is varied from 1 to 3 with the step size of 1.53 dB. scaling the rotated image.96 0. sharpening.43 0. The proposed method can effectively resist attacks such as median ﬁlter with a mask of size 5 Â 5. 60. average ﬁlters. and (c) Peppers.90 (5 Â 5) 0. average ﬁltering attacks (3 Â 3. (a) Lena.99 1.95 0. Gaussian ﬁlter.34 15 0.97 80 0.68 0. For geometric attacks.99 0. The proposed method can detect the existence of a watermark through quality factors greater than 20. Lin et al. it will cause errors in the watermark extraction process.86 0.67. and histogram equalization.99 90 0.84 0. The attack of scaling is performed by ﬁrst shrinking an image of pixels 512 Â 512 to 256 Â 256 via PhotoImpact 11 software.75 0.96 0. for the cropping attack.55 20 0.97 0.82 35 0.93 0. The results are shown in Table 3.91 0. The rotational attacks are performed by rotating an image at a small angle.90 50 0.99 0. (b) Goldhill.76 0. Neogeometric attacks Median ﬁlter (3 Â 3) (a) Lenna NC EXtracted watermark (b) Goldhill NC EXtracted watermark (c) Peppers NC EXtracted watermark 0. The results show that the value of NC is still greater than 0. the image compression is improved. 1997). and (c) Peppers. 50.81 0. our method’s PSNR is higher than Lien’s. 90. and cropping the scaled image to the original image size (Petitcolas.97 0. 1997) are applied in the experiments. The resulting extracted watermark is shown in Table 3.96 0. Although Lien improves on Wang and Lin’s (2004) approach to watermarked image quality.’s (2006).92 0.94 0.96 60 0. 5 Â 5.1. and each cropped pixel is replaced with 255. The JPEG quality factor is a number between 0 and 100 and associates a numerical value with a particular compression level. For a pixel-based watermarking system.

Lausanne.51 NA 0. B. M.99 0. R.67 0. IEEE Transactions on Industrial Electron. Horng. Inoue.79 Proposed method (PSNR = 42. et al.46 0.84 0. Hiding digital watermarks using multiresolution wavelet transform.51 0. 48(5).25°) Rotation (degree: À0. (2001).59 0.-S.85 0. Optimal differential energy watermarking of DCT encoded images and video. Leighton. Hsieh.54 0. R. pp. Geometric attacks Scaling Cropping Gaussian noise Variance 1 (a) Lenna NC Extracted watermark (b) Goldhill NC Extracted watermark (c) Peppers NC Extracted watermark 0. As a result. Gaussian noise added by variations from 1 to 3.. Comparison of two wavelet based image watermarking schemes. C. average ﬁlter 5 Â 5.70 0.54 dB) 0.3 0. J. a blind watermarking method based on the DWT using maximum wavelet coefﬁcient quantization is proposed. N.65 0. Man and Cybernetics. and (c) Peppers.. G.89 0. Image digital watermarking algorithm using multiresolution wavelet transform.55 0. Cox. A.25 0. References Chen. I. In International Multimedia Conference on Multimedia and Security. M. Magdeburg. cropping 1/4. In Proceedings of the IEEE AVSS. Highly reliable digital watermarking using successive subband quantization and human visual system. Huang. H..61 0. (2001). the proposed scheme can also be applied to data hiding and image authentication.59 Degree 0. IEEE Transactions on Image Processing.15 0.59 0. 1221–1229. Kilian.. Chicago. Conclusion In this paper..51 0.26 NA 0. H. Kwon.32 NA NA Li et al. W. 682–685. & Lagendijk.88 0.61 0..79 0. Bouridane. pp.. pp. 419–423. F. À: counter clockwise). Kilian. (1997).82 0. and rotation ﬂow by scaling and cropping to the original size of the three test images. et al.. & Yang. Ratakonda.. et al. L. C..78 0. S. PrenticeHall. in the proposed method the watermarked images look lossless in comparison to the original images and the watermark can effectively resist common image processing attacks. Digital image processing (2nd ed. 391–395.97 0. and Gaussian noise with a variation of less than 2. (b) Goldhill.23 NA NA NA 0.. 2977–2982. & Eskicioglu.64 0. (2005).. Chicago. & Woods. Kobe. D. Robust DWT-SVD domain image watermarking: embedding data in all frequencies.79 0. A variable block-size dot-based watermarking method. S. & Huang. and Lien and Lin’s (2006) methods.45 Degree À0.35 Lien and Lin (2006) (PSNR = 41.44 0.32 0. Miyazaki.66 0.62 Variance 3 0. J.88 0. Y. the proposed method is more robust in resisting common attacks such as median ﬁlter 5 Â 5. (2005).. R.17 0. A robust wavelet-based digital watermarking using level-adaptive thresholding. 148–158. K. IL.34 Rotation Degree 0.35 0..53 0. Leighton. F.76 0.54 0.-H.. A. A. G.90 0. F. (1996). IEEE Transactions on Industrial Electronics. 226–230.46 0. 5.61 0. 52(1). et al. J. In Proceedings of IEEE ISIE.38 0..88 0. In Proceedings of the IEEE ICIP. / Expert Systems with Applications 36 (2009) 11509–11516 11515 Table 3 Normalized correlation coefﬁcients (NC) after attacked by scaling 256 Â 256.64 0. Kim. Ha.47 0.65 0.50 0. (2004).34 0. & Park.. Attacks n NC Median ﬁlter (3 Â 3) Median ﬁlter (4 Â 4) Median ﬁlter (7 Â 7) JPEG (QF = 10) JPEG (QF = 20) JPEG (QF = 30) JPEG (QF = 50) JPEG (QF = 70) JPEG (QF = 90) Sharpening Gaussian ﬁlter Rotation (degree: 0. In Proceedings of the IEEE International Conference on Systems. IEEE Transactions on Consumer Electronics. In Proceedings of the IEEE ICIP. 875–882. Li et al. (2002). 2005.. (1998)...81 Variance 2 0. pp. pp.25°) Cropping 1/4 Scaling 256 Â 256 Wang and Lin (2004) (PSNR = 38. Kim. J. W.63 0. J.90 0. A publicly veriﬁable copyright-proving scheme resistant to malicious attacks. Cox. Lin et al. Davoine. I. In addition to copyright protection.. Table 4 Comparing the proposed method with Wang and Lin’s (2004). Yamamoto. Langelaar. G.. A. Como.02 B) 0.37 0.60 Degree À0. (2001). & Ahuja. audio and video. (1998).46 0. H. Tseng. Moon. Kwon. C. 205–209.66 0.88 0. pp. & Lee. pp.. J. R. T. Ganic. Acknowledgement This work was partially supported by National Science Council of Taiwan under the Contract Numbers NSC-97-2221-E-129-007 and NSC-97-2221-E-239-022. Secure spread spectrum watermarking for multimedia.. E. Gonzalez. 1673–1687. The proposed method is different from the block-based or wavelet-tree based methods which are discussed in Section 1.44 0. I. et al. 6(12). 588–592. E.97 1 0. (2006) (PSNR = 40. In Proceedings of the IEEE ICIP. T. Pusan. In International Conference on Image Processing. Secure spread spectrum watermarking for images. .84 0. we use variable blocks size and embed a watermark bit using different subbands which are selected randomly from two subbands.96 0. F. pp. R.57 1 0. S.15 0.. Kurugollu. (2000). Vancouver. 45(4). IEEE Transactions on Image Processing. H.52 0.92 0.34 0. A digital watermark based on the wavelet transform and its robustness on image compression. especially by JPEG compression (with a quality factor greater than 20) and low-pass ﬁlter. R. A new wavelet-based scheme for watermarking images.60 0.97 0.).52 0. (1999).W. pp.. Ban.52 0. Dugad.88 Unit of rotationt: degree (+: clockwise. Kheliﬁ.55 (a) Lena. 243– 246. 166–174. In order to achieve the secrecy of embedding watermark.53 0. T.2 dB) 0. S. (1999).’s (2006).6 dB) 0. An improved wavelet-based image watermarking technique.. Moreover. O.28 0. (2004).25 0. 327–334. In IEEE ICIP. 10(1). Y.3 0.. Y.

In Proceedings of the IEEE INDIN.-H. pp. (2000). In Thirty-Fourth Asilomar Conference on Signals. C. A. & Kuo. IEEE Transactions on Signal Processing. B. Schilling.. 13(2). Graphics and Image Processing. 855–884. 383–392. H. J. (1997). (2000).. 623– 626. In Proceedings of the IEEE ICIP. Huo. F. K. J.. Tsai. In Proceedings of the IEEE ICIP. Han. J. Wang. Lin et al. Wavelet packet and adaptive spatial transformation of watermark for digital images authentication. 1350–1354. C. (2007). Zhang. pp. pp. C. pp. C. Petitcolas. Temi. L. & Kuo. Systems and Computers (pp.. Available from: <http://www. 1846–1850). (1993). In Proceedings of the IEEE WCICA. & Lin.petitcolas. (2007). & Milstein.. Q. Wang. Lien. IEEE Transactions on Image Processing. Y. P. 41(12). A watermarking method based on maximum distance wavelet tree quantization. Wang.. Potdar.. A.. 709–716. Wang. 505–516. J. K. High ﬁdelity image compression with multithreshold wavelet coding (MTWC). Vancouver.. (2005). 10256–10260.. 3(12).. V. C. H. / Expert Systems with Applications 36 (2009) 11509–11516 Tsai. H.. Inc. 491–496. & Uhl. pp. A multi-threshold wavelet coder (MTWC) for high ﬁdelity image compression. (1982). P. Y. H.. Santa Barbara. In Proceedings of the IEEE ISCIT. Wavelet-based digital image watermarking. Wu. Speech.. S. Embedded image coding using zerotrees of wavelet coefﬁcients. S. & Wen. J. An integer wavelet based multiple logowatermarking scheme. An adaptive block-based blind watermarking algorithm.net/fabien/watermarking/stirmark/index. 315–319. Signal Processing]. Electronic Imaging. J. Y. 269–276. S. In ICIS. J. 154– 165.. Li. M. E. Wavelet tree quantization for copyright protection watermarking. & Lin. & Cheng. D. P. In Proceedings of the IEEE ICC Glasgow. A. P. Wang. & Chang. (1997). 450–453. Optics Express. Weakness of existing watermark scheme. Image watermarking using structure based wavelet tree quantization. M. C. P. In SPIE’s Annual Meeting-Application of Digital Image Processing XX. 652–655. M. (2001). Scotland. Chen. Theory of spread-spectrum communications – A tutorial. H. P. In Proceedings of the SPIE. Security and Watermarking of Multimedia Contents III. & Lasakul. M. Yu. pp. pp.. Constrained wavelet tree quantization for image watermarking. R. Ulead Systems. 3445–3462 [see also IEEE Transactions on Acoustics. (1998). L. C. Shapiro. J. A robust image watermarking using multiresolution analysis of wavelet. A survey of digital image watermarking techniques. pp. Choomchuay. W. . Robust image watermark with wavelet transform and spread spectrum techniques. (2006). Wang. pp. Meerwald. Y. pp. Liang. X. H. M. 1997. BC. In Proceedings of the IEEE ICSP. S. G. PhotoImpact 11 software. H. & Niu. & Huang.. & Chen.. pp. 2294–2297. (2006). Z. In 19th Conference on Computer Vision. Australia. D. Melbourne.. Paciﬁc Grove.. (1997). J. C. (2005). A survey of watermark-domain watermarking algorithms.html>. (2004). J. E. & Lin. IEEE Transactions on Communication. C.... Y. Su. G.. 30(5). (2004). Pickholtz. P. San Diego.11516 W.

- 13042011_Lezione_Interpolazione
- 27962
- 24504
- 22-00651210
- 10.1.1.70.359
- 7_13_eng_2
- 7_13_eng
- 2
- Secure Transmission of Medical Image by Blind Watermarking Technique
- A Session Based Blind Watermarking Technique within the NROI of Retinal Fundus Images for Authentication Using DWT, Spread Spectrum and Harris Corner Detection
- A BLIND WATERMARKING TECHNIQUE IN JPEG COMPRESSED DOMAIN
- A New Wavelet Based Scheme for Blind Watermarking
- Verbal and Nonverbal Communication:
- 63417751 MCQ Indian History

This paper proposes a blind watermarking algorithm based on maximum wavelet coefficient quantization
for copyright protection. The wavelet coefficients are grouped into different block size and bl...

This paper proposes a blind watermarking algorithm based on maximum wavelet coefficient quantization

for copyright protection. The wavelet coefficients are grouped into different block size and blocks

are randomly selected from different subbands. We add different energies to the maximum wavelet

coefficient under the constraint that the maximum wavelet coefficient is always maximum in a block.

The watermark is embedded the local maximum coefficient which can effectively resist attacks. Also,

using the block-based watermarking, we can extract the watermark without using the original image

or watermark. Experimental results show that the proposed method is quite robust under either

non-geometry or geometry attacks

for copyright protection. The wavelet coefficients are grouped into different block size and blocks

are randomly selected from different subbands. We add different energies to the maximum wavelet

coefficient under the constraint that the maximum wavelet coefficient is always maximum in a block.

The watermark is embedded the local maximum coefficient which can effectively resist attacks. Also,

using the block-based watermarking, we can extract the watermark without using the original image

or watermark. Experimental results show that the proposed method is quite robust under either

non-geometry or geometry attacks

- VLSI Architecture of Invisible Watermarking Using Lsb and Lifting Scheme for Real Time Copyright Protection
- Iris Pattern Recognition Using Complex Wavelet And Wavelet Packet Transform
- ijatcse02132912
- MSing
- Embedded Zerotree Wavelet
- Ch_15_4 EZW (PPT) revised
- Image Adaptive Watermarking Using Wavelet Domain Singular Value
- 539-566
- SPIHTJournal
- ROBUST WATERMARKING TECHNIQUES
- PERFORMANCE COMPARISON OF HYBRID WAVELET TRANSFORMS FORMED USING DCT, WALSH, HAAR AND DKT IN WATERMARKING
- Report
- A DWT-based Image Watermarking Approach Using Quantization on Filtered Blocks
- dwt
- Kiran Reference 4
- Csc461 Monia Wavelet
- 100310
- 631F01_lec9
- Embedded+Image+Coding+Using+Zerotrees
- Welcome to International Journal of Engineering Research and Development (IJERD)
- Wave Let
- Wavelet Based Watermarking
- pxc3876666
- Robust Image Watermarking Using Non-Regular Wavelets
- Dr 24751755
- Adaptive Directional Wavelet Transform Based on Directional Prefiltering-DpM
- wavelet_mmsp05
- 31295019150027
- International Journal of Engineering Research and Development (IJERD)
- Analysis of Multi Resolution Image Denoising Scheme using Fractal Transform
- A blind watermarking method using maximum wavelet coefficient quantization

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd