Jia-Hong Lee
National Kaohsiung First University of Science and Technology
Department of Information Management Kaohsiung, Taiwan, R.O.C.
Yu-Kuen Ho
National Cheng Kung University Department of Electrical Engineering Tainan, Taiwan, R.O.C.
like scheme is designed to embed and extract invisible watermarks. In contrast with the previous object-based watermarking schemes, extra information for resynchronization is not required to be stored in our method. Experimental results show the robustness of the proposed method against many kinds of geometrical attacks. © 2007 Society of Photo-Optical Instrumentation Engineers. 关DOI: 10.1117/1.2751167兴
Subject terms: object-based watermarking; shape subdivision; principal axis; geometrical attacks; resynchronization detection problem; false-positive probability.
Paper 060367RR received May 13, 2006; revised manuscript received Dec. 20, 2006; accepted for publication Jan. 5, 2007; published online Jul. 2, 2007.
1 Introduction
The rapid growth of commercial multimedia services has led to an urgent demand for reliable and secure copyright protection for digital multimedia. Digital watermarking techniques were first introduced in the early 1990s, and are being rapidly developed for various media. In many water-marking applications, watermarks have to survive some ex-pected set of distortions, including digital filtering, lossy compression, analog blurring, addition of random noise, and ghosting. Many robust watermarking techniques, in-cluding statistics,1 signal transformation,2 spread spectrum,3 the discrete cosine transform 共DCT兲,4 the dis-crete Fourier transform 共DFT兲,5 wavelets,6 the Fourier-Mellin transform,7 fractals,8 and content-based methods,9 can be efficiently adopted to add watermarks into digital images. The stego images generated using these schemes can survive common image processing, such as lossy com-pression, filtering, noise addition, and geometrical transfor-mations. Cox et al.10 give a comprehensive description of watermarking techniques.
However, geometrical processes like simple rotation, scaling, and translation 共RST兲 are easily performed using certain commercial software such as PhotoImpact and Pho-toShop. For this reason, some methods have been proposed to resist RST distortions. Wu et al.11and Lin et al.12 adopt the Fourier-Mellin transform to achieve geometrical-transform invariance. These methods take the whole image into account in embedding and detection processes, though the demand may be for protecting a single object in the image.
Object-based watermarking schemes13,14 have become more important with the quick developing of the internet and the multimedia standards. Objects in a video frame or
pictures on auction Web sites with simple background can easily be separated from their background by using the bi-nary mask provided by the MPEG-4 video format or image-processing tools. Watermark schemes that embed watermarks into a whole image without considering that the image could be misappropriated by removing the back-ground could fail in watermark detection or extraction. Be-sides, geometrical processes like rotation and skewing could be applied to the usurped object image, making the detection or extraction process more difficult. With crop-ping and geometrical attacks, watermarks embedded in the media retain most of their original data, but lose the origi-nal order and orientation of the embedded data, which are critical for watermark extraction or detection. This problem is called the watermark resynchronization detection prob-lem.
With the prevalence of image-processing tools, the re-synchronization detection problem is getting more impor-tant for the practical use of watermark schemes. Guo and Shi15 and Lu and Liao16 recently adopted inertia ellipses and eigenvectors to correct object distortions resulting from geometrical attacks such as rotation and scaling. However, in the method proposed by Guo and Shi, information about the inertia ellipse of the object before watermarking is re-quired for watermark detection. This means that extra in-formation about the inertia ellipse of the original object needs to be recorded for resynchronization. Without the information about the original inertia ellipse, the object cannot be resynchronized to the original dimensions and angular magnitude after the scaling and rotation attacks.
The same situation also happens in the method proposed by Lu and Liao.16 It is a video object watermarking scheme, based on the concept of communications with side information. Using the eigenvectors of a video object, Lu and Liao’s method is able to solve the resynchronization problems caused by geometrical attacks like rotation and
flipping. However, embedding watermarks using the DCT coefficients will make this method very sensitive to such attacks. Therefore, the original dimensions and eigenvec-tors are also needed as extra information in watermark de-tection for resynchronization.
In our previous work,17 we have proposed a spatial-domain method based on shape self-similarity segmentation to achieve synchronization detection. This method does not have to record extra information for resynchronization. However, shape self-similarity segmentation is good for convex objects only; it is not general enough for all kinds of object shapes.
To make it general enough for practical applications, we propose an object-based watermarking scheme using a new segmentation algorithm for object images. The proposed method is based on the orientations and lengths of principal axes, and is rotation- and translation-invariant. As long as the principal axes of an object are detected, watermarks embedded in segmented regions can be extracted correctly. In addition, to ensure the reliability of the proposed method, the issue of false positives is discussed in this pa-per. Figure 1 shows the procedure of the proposed scheme. The rest of this paper is organized as follows. Section 2 briefly describes the proposed method of segmenting an object image into numerous small subregions. Section 3 then details the procedures of watermark embedding and extraction. Next, Sec. 4 describes experimental results con-cerning some attacks, including the mean filter, JPEG com-pression, rotation, rotation and scaling, and geometrical at-tacks, using the StirMark18,19 benchmark tool. We also implemented Lu and Liao’s16method for comparison. Sec-tion 5 discusses in detail the experimental results of imple-mented methods. Conclusions are finally drawn in Sec. 6.
2 The Proposed Shape Subdivision Method for Synchronization
Let f denote an object image segmented from a gray-level or color image, whose corresponding shape S is then ob-tained in binary image form by transforming all pixel illu-minations of f to the same gray value. The shape subdivi-sion corresponding to the principal axes is performed to ensure the success of embedding and extracting operations. This procedure is described as follows:
Step 1: Determine the second-order central moments of S, namely,02,20, and 11.
Step 2: Determine the principal angle of S from the
central moments as =1 2 tan −1
冉
211 20−02冊
. 共1兲Step 3: Determine the minimum bounding rectangle of the rotated object S with rotation around its mass center c at an angle of , so that the length and width of the rectangle are equal to the lengths of computed principal axis and subprincipal axis, respectively.
Step 4: Divide the object image by segmenting the bounded rectangle into m⫻n small similar rectangles. An experiment was conducted using an object image called “Russian Doll” The second-order central moments
02, 20, and 11 of this object image are calculated as 1.36⫻108, 2.77⫻108, and −2.457⫻107, respectively. Us-ing these second central moments and Eq.共1兲, the principal angle can be determined. Figure 3a shows the original object image, whose principal angle is obtained as = 1.29 deg. The dimensions of its minimum bounding rect-angle are 300⫻600.
3 Watermark Embedding and Extraction
Suppose that an object image f is divided into m⫻n dis-joint gray-level image blocks, B1
⬘
, B2⬘
, . . . , Bm⬘
⫻n, using the segmentation method mentioned in the preceding section. The parameters m and n can be regarded as secret keys to embed and protect data. All blocks Biare then split into k groups, G1, G2, . . . , Gk, where k denotes the number of bits in the watermark to be embedded. Each group Gi can be denoted as Gi=再
Bj:共i − 1兲 ⫻ m⫻ n k + 1艋 j 艋 i ⫻ m⫻ n k , j苸 N冎
. 共2兲 Furthermore, Gi is divided into two subsets Ci and Di, which contain all the blocks Bj in Gi with odd and even indices, respectively.Finally, the following is applied to embed the watermark bit wi:
兩C¯i− D¯i兩 艌␦u if wi= 1,
the effect of blockness. The gradient of each pixel com-puted here is given by
gp= max
q苸D共兩p − q兩兲, 共4兲
where D denotes an area centered on pixel p. To avoid the effect of blockness, we adjust the intensity of each pixel according to the capacity of modification, which is given by Cp=
再
log2共gp兲 if gp⬎ 0,
0 otherwise.
冎
共5兲This equation guarantees that a pixel will not be modified if its gradient is very small. Moreover, to blur the boundaries of blocks, 2-D Gaussianlike shape weights can be applied in pixel modification. Equation共5兲 can be replaced by
Cp=log2共gp兲 ⫻ ⫻ Gau គ wp, 共6兲
Where denotes the strength of the signal, and Gauគwpis the corresponding weight of pixel p, which is expressed as Gauគ wp共x,y兲= exp关− d共x,y兲/22兴. 共7兲 Here d共x,y兲 is the distance between the pixel p共x,y兲 and the center of the block in which it is located.
Figure 2 shows an example of the use of these equations to embed a 32-bit watermark into the object image seg-mented from “Lena.” Figure 2共a兲 is the original object im-age, Fig. 2共b兲 is the watermarked imim-age, and Fig. 2共c兲 shows the 2-D Gaussian shape weights used. Figure 2共b兲, with high resolution, shows the imperceptibility of pro-posed method.
To extract watermarks from a test object, the normaliza-tion and segmentanormaliza-tion procedures performed in the water-mark embedding process are performed again with the same specified secret keys. The corresponding blocks can then be calculated, and watermarks can be extracted easily by using Ciand Di. If兩Ci− Di兩艌共␦u+␦l兲/2, then the bit 1 is extracted; otherwise the bit 0 is extracted.
For example, to embed 32 bits into 24⫻24 blocks, the blocks are split into 18 groups. Each group contains 32 blocks. Then, each group is divided into two subsets for watermark embedding. To embed the watermark, the di-mensions of D in Eq.共4兲 are set to be 2⫻1, the value of in Eq.共5兲 is 420, and the value of␦, the standard deviation of the 2-D Gaussian, in Eq.共6兲 is 5.0. The difference of the gray-level values of the two subsets is adjusted to be bigger than a threshold␦u共␦u= 8 is used in the experiment shown in Fig. 3兲 if a watermark bit 1 is embedded. On the other hand, the difference is adjusted to be smaller than ␦l 共␦l
Fig. 2 Proposed scheme applied to a high-resolution object image
“Lena.”共a兲 The test object image. 共b兲 The object image in 共a兲 water-marked using the proposed scheme.共c兲 The 2-D Gaussian weights amplified 20 times for display.
= 2 is used兲 if a 0 is embedded. Figure 3共b兲 illustrates the shape segmentation result of dividing Fig. 3共a兲 into 24 ⫻24 small blocks in which 32 watermark bits were embed-ded. At the stage of watermark extraction, a bit 1 is ex-tracted if the difference of the sum of gray-level values of two subsets is larger than共␦u+␦l兲/2, and a bit 0 is extracted if the difference is smaller than共␦u+␦l兲/2.
4 Experimental Results
Experiments were conduced using the object images “Rus-sian Doll,” “Vase,” “Fish,” and “Sea Shell.” Some of these object images are neither convex nor symmetric. To evalu-ate the proposed scheme, two similarity measures, PSNR and NC, are used. The PSNR共peak SNR兲 is used to mea-sure the gray-image quality, and the NC共normalized corre-lation兲 is used to measure the similarity between two bi-level watermarks. The NC20 is defined as follows:
NC =
兺
i wiwi⬘
兺
i wi2兺
i 共1 − wi兲共1 − wi⬘
兲兺
i 共1 − wi兲2 , 共8兲where w and w
⬘
are two watermark sequences. The NC value is the proportion of the number of watermark bits that have the same value at the same location in both watermark sequences. Different attacks were simulated in these experi-ments, including mean filtering, JPEG compression, noise, resizing, rotation, and row and column removal.To evaluate the proposed method, the method proposed by Lu and Liao is implemented for comparison. Their scheme works on an object image in a MPEG-4 video frame, which can be segmented from its background using a binary mask. The scheme can also be applied to embed watermarks into an ordinary still image as long as the ob-ject image can be segmented from its background accu-rately every time we want to process it. Pictures that have simple backgrounds, so that the object can be easily sepa-rated from its background共as in most merchandise images found on auction Web sites兲 are quite proper for this usage. 共To make the object more conspicuous to the viewers, pic-tures on an auction Web site are likely to have a simple
background.兲 The proposed method, like the method pro-posed by Lu and Liao, can also be used to process object images in an MPEG-4 digital video as well as the object image in an ordinary still image with a simple background. Since the method proposed by Lu and Liao uses water-mark detection, for a proper comparison a waterwater-mark de-tection experiment has been designed. For Lu and Liao’s method, 10,000 sets of Gaussian-distributed watermark se-quences with zero mean and unit standard deviation are generated. The correlations between the sequences and the watermark extracted from the watermarked image are cal-culated. For the proposed method, 10,000 sets of randomly generated watermark sequences with equal numbers of 0’s and 1’s are used to calculate the NC between the random sequence and the watermark extracted from the water-marked image.
Figures 3共a兲, 4共a兲, 5共a兲, and 6共a兲 show the original object images. The principal angles were obtained as 1.29, 0.1, −11.83, and −7.15 deg for “Russian Doll,” “Vase,” “Fish,” and “Sea Shell,” respectively. Figures 3共b兲, 4共b兲, 5共b兲, and 6共b兲 show the shape segmentation results of dividing origi-nal object images into 24⫻24 small blocks in which 32 watermark bits were going to be embedded. Figures 3共c兲, 4共c兲, 5共c兲, and 6共c兲 show the corresponding detector re-sponse using the proposed method. Figures 3共d兲, 4共d兲, 5共d兲, and 6共d兲 show the corresponding detector response using Lu’s method. The 500th of the 10,000 sequences is re-placed with the embedded watermark. From these figures, only one peak exists共at the position of the 500th兲, and an acceptable watermark detection threshold Tz is adaptively determined using the following equation:
Ts=+ k, 共9兲
whereand are the mean value and standard deviation of the 10,000 NC values, calculated by correlating the wa-termark with 10,000 random sequences, respectively. In our experiments, the of Lu and Liao’s and the proposed method are about 0 and 0.25, respectively, and the are about 0.07 and 0.1, respectively. Assume that these 10,000 NC values are normally distributed. Let f共x兲 be a normal probability density function. To make the probability of
Fig. 3 Proposed scheme applied to the “Russian Doll” object.共a兲 The test object image. 共b兲 The shape
subdivision of the image in共a兲. 共c兲 The corresponding detector response using the proposed method. 共d兲 The corresponding detector response using Lu and Liao’s method.
false alarm less than , the value of Tz must satisfy the following inequality:
冕
Tz ⬁ f共x兲dx 艋, 共10兲 where f共x兲 =冑
1 2 exp冋
− 共x −兲2 22册
.In our experiments, the value 4.265 of k is determined to make the probability of a false positive of a random
water-mark less than 10−5. The threshold values of Tzfor Lu and Liao’s method and the proposed method are 0.3 and 0.68, respectively. The experimental results of several kinds of attacks using these two methods are shown in Table 1. A study of the threshold selection for correlation can be found in Ref. 21.
5 Discussion
Digital watermarking algorithms in the frequency domain are well known to resist general image-processing attacks better than those in the spatial domain. The experimental results also show that Lu and Liao’s method performed
Fig. 4 Proposed scheme applied to the “Vase” object.共a兲 The test object image. 共b兲 The shape
subdivision of the image in共a兲. 共c兲 The corresponding detector response using the proposed method. 共d兲 The corresponding detector response using Lu and Liao’s method.
Fig. 5 Proposed scheme applied to the “Fish” object.共a兲 The test object image. 共b兲 The shape
sub-division of the image in共a兲. 共c兲 The corresponding detector response using the proposed method. 共d兲 The corresponding detector response using Lu and Liao’s method.
slightly better than the proposed method in some image-processing attacks, such as mean filtering and uniform noise addition. However, the proposed method performed much better than that of Lu and Liao in resisting general
geometrical attacks such as row or column removal. In ad-dition, we assume that the parameters of the affine transfor-mation 共rotation and scaling兲 attacks in Lu and Liao’s method are known, so we can reconstruct the original shape
Fig. 6 Proposed scheme applied to the “Sea Shell” object.共a兲 The test object image. 共b兲 The shape
subdivision of the image in共a兲. 共c兲 The corresponding detector response using the proposed method. 共d兲 The corresponding detector response using Lu and Liao’s method.
Table 1 Experimental results on robustness against attacks.
Target object image:
NC
“Russian Doll” “Vase” “Fish” “Sea Shell”
Method: Ours Ref. 16 Ours Ref. 16 Ours Ref. 16 Ours Ref. 16
PSNR: 38.20 41.39 39.26 39.55 39.96 39.37 38.02 43.31
Watermarked共no attack兲 1.00 0.80 1.00 0.64 1.00 0.59 1.00 0.69
Mean filter 3⫻3 0.88 0.80 0.94 0.66 0.82 0.63 1.00 0.69 5⫻5 0.88 0.80 0.94 0.68 0.82 0.63 0.71 0.70 IPEG共20%兲 0.94 0.58 0.94 0.16a 0.72 0.57 0.94 0.02a Uniform = 10 1.00 0.80 1.00 0.64 1.00 0.59 1.00 0.68 = 15 0.94 0.80 0.35a 0.64 0.88 0.59 1.00 0.68 = 20 0.94 0.80 0.18a 0.64 0.35a 0.58 0.94 0.67 Rotation共20 deg.兲 1.0 0.72 0.92 0.62 0.82 0.59 1.00 0.14a Scaling: 0.8⫻0.8 0.88 0.72 0.88 0.64 0.82 0.59 0.94 0.69 0.9⫻0.9 0.94 0.72 0.94 0.62 0.88 0.59 1.00 0.69 1.1⫻1.1 0.88 0.72 1.0 0.62 0.72 0.59 0.94 0.69
One row or column removal 1.00 0.20a 1.0 0.45 0.94 0.11a 1.00 0.15a
Two rows or columns removal 0.92 0.05a 1.0 0.21a 0.88 0.09a 0.88 0.14a
Fig. 7 Distortion results using StirMark and their corresponding BER and NC values by using the
proposed method. “Russian Doll:”共a兲 affine transform 共BER=3/32; NC=0.82兲; 共b兲 random distortion 共 BER= 5 / 32; NC= 0.71兲; 共c兲 line removal 共BER=3/32; NC=0.82兲. “Sea Shell:” 共d兲 affine transform 共 BER= 3 / 32; NC= 0.82兲; 共e兲 random distortion 共BER=2/32; NC=0.88兲; 共f兲 line removal 共BER=0; NC = 1兲. “Fish:” 共g兲 affine transform 共BER=4/32; NC=0.76兲; 共h兲 random distortion 共BER=5/32; NC = 0.70兲; 共i兲 line removal 共BER=2/32; NC=0.88兲. “Vase:” 共j兲 affine transform 共BER=5/32; NC=0.71兲; 共k兲 random distortion共BER=5/32; NC=0.71兲; 共l兲 line removal 共BER=3/32; NC=0.82兲.
before performing watermark detection. Therefore, the ex-perimental results seem to be good in these cases. In prac-tical applications, however, Lu and Liao’s method will suf-fer from the resynchronization problem, since those parameters are in fact unknown.
Furthermore, certain commercial software applications, including PhotoImpact and PhotoShop, provide a boundary-pixel cropping function to erode the boundary pixels of object images. Row or column removal and boundary-pixel cropping attacks will change the image di-mensions. Without the information of the original image dimensions, any DCT-based method will fail in watermark detection. Besides, after some geometrical attacks such as affine transformation and random distortion, since the DCT coefficients are sensitive to geometrical distortion, the em-bedded watermark is still difficult to detect even if the di-mensions of the original image have been recorded.
To evaluate the effectiveness of the proposed method, some geometrically distorted results generated by the Stir-Mark benchmark tool are used for testing. These distortions include affine transformation, random distortion, and the removal of a large number of lines. Experimental results illustrated in Fig. 7 showed that the proposed method is quite effective in resisting these attacks. The bit error rate 共BER兲 was used to measure the similarity between the K-bit pattern v =共v0,v1, . . . ,vk−1兲 and the original water-mark w:
BER =兩兵vj:vj⫽ wj其兩
K .
The NC values shown in Table 1 are all larger than the thresholds Tzof 0.68. These results show our method to be successful in watermark detection.
We also applied these attacks to Lu and Liao’s method. Since the DCT-based methods are quite sensitive to geo-metrical distortion, without knowing the degree of distor-tion as side informadistor-tion, the NC values obtained using Lu and Liao’s method are all in the range 共−0.1,0.1兲 and are undetectable.
Segmenting an object from an image could be fatal to watermark detection using a non-object-based watermark-ing approach.22 The proposed object-based watermarking scheme can select relevant objects separately and embed watermarks into the selected objects. These objects should be the most important components of the image. When a certain region containing one of these objects is cut from the image and poached, the watermarks should still be de-tected accurately with the proposed method. Figure 8 shows an example of embedding watermarks into a selected object, namely the face “Lena.” The experimental result indicates that a high NC value of 0.82 is obtained even following rotation, scaling, and boundary-pixel cropping at-tacks.
Moreover, the proposed scheme can efficiently achieve the synchronization recovery that is required for recovering from some geometrical attacks, such as skew, when per-forming simple operations in shape subdivision. The skew operation changes the locations of the principal axes of image objects. Figure 9 shows an example of synchroniza-tion recovery from skewing the image “Vase.” The x and y axes in Fig. 9共c兲 represent the orientation adjustment of the
two principal axes. The output values on the z axis are the NC values of watermark detection. The maximum NC value indicates the accurate angular magnitude of skewing. The execution performance of the proposed algorithm is very good. It only has a time complexity of O共M ⫻N兲 to embed or extract watermark bits in an image with size M ⫻N, since the major operations, including finding principal axes, pixel grouping, and gradient computation, all have O共M ⫻N兲 time complexity. The execution times of the pro-posed method for an image of size 300⫻600 on a PC with 3-GHz CPU and 1-Gbit memory are 0.18 and 0.11 s for watermark embedding and extraction, respectively.
6 Conclusions
This study has presented a new object-based watermarking method, based on efficient segmentation, using parallel and perpendicular lines, of two principal axes in the spatial do-main. This proposed scheme also was compared with the DCT-based method developed by Lu and Liao. Four object images taken from product images on the eBay auction Web site were adopted for model testing. Experimental re-sults indicate that the proposed scheme resists geometrical attacks more effectively than that of Lu and Liao. The pro-posed method was also applied to an object image cut from
Fig. 8 An interactive object watermarking application using the
pro-posed scheme.
Fig. 9 An experiment on skew calibration by searching for the
nearby angles of computed principal-axis orientations. 共a兲 The skewed “Vase” image.共b兲 The principal-axis detection for the object in共a兲. 共c兲 Synchronization recovery by searching for correct orienta-tions of the two principal axes.共d兲 The principal axes after recovery.
References
1. W. Bender, D. Gruhl, N. Morimoto, and A. Lu, “Techniques for data hiding,” IBM Syst. J. 35共3&4兲, 313–336 共1996兲.
2. C.-T. Hsu and J.-L. Wu, “Hidden signatures in images,” in Proc. Int.
Conf. on Image Processing, pp. 223–226, IEEE共1996兲.
3. L. M. Marvel, C. G. Boncelet, and C. T. Retter, “Spread spectrum image steganography,” IEEE Trans. Image Process. 8共8兲, 1075–1083 共1999兲.
4. I. J. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, “Secure spread spectrum watermarking for multimedia,” IEEE Trans. Image Process.
6, 1673–1687共1997兲.
5. V. Solachidis and I. Pitas, “Circularly symmetric watermark embed-ding in 2-D DFT domain,” IEEE Trans. Image Process. 10共11兲, 1741–1753共2001兲.
6. H. Inoue, A. Miyazaki, A. Yamamoto, and T. Katsura, “A digital watermark based on wavelet transform and its robustness on image processing,” in Proc. Int. Conf. on Image Processing, Vol. 2, pp. 391–395, IEEE共1998兲.
7. P. Bas, J.-M. Chassery, and F. Davoine, “Using the fractal code to watermark images,” in Proc. Int. Conf. on Image Processing, Vol. 1, pp. 469–473, IEEE共1998兲.
8. V. Solachidis and F. M. Boland, “Phase watermarking of digital im-ages,” in Proc. Int. Conf. on Image Processing, Vol. 3, pp. 239–242, IEEE共1996兲.
9. P. Bas, J.-M. Chassery, and B. Macq, “Geometrically invariant
wa-15. H. Guo and P. F. Shi, “Object-based watermarking scheme robust to object manipulations,” Electron. Lett. 38共25兲, 1656–1657 共2002兲. 16. C. S. Lu and H. Y. M. Liao, “Video object-based watermarking: a
rotation and flipping resilient scheme,” in Proc. Int. Conf. on Image
Processing, Vol. 2, pp. 483–486共2001兲.
17. M. Y. Wu and Y. K. Ho, “A robust object-based watermarking scheme based on shape self-similarity segmentation,” in 5th Int.
Symp. on Multimedia Software Engineering, pp. 110–113, IEEE
共2003兲.
18. F. A. P. Petitcolas, R. J. Anderson, and M. G. Kuhn, “Attacks on copyright marking systems,” in Information Hiding, Second Int.
Workshop, IH’98, Proc., pp. 219–239, Springer-Verlag共1998兲.
19. F. A. P. Petitcolas, “Watermarking schemes evaluation,” IEEE Signal
Process. Mag. 17共5兲, 58–64 共2000兲.
20. P. M. Chen, “Robust digital watermarking based on a statistic ap-proach,” in 2000 Int. Symp. on Information Technology (ITCC2000), pp. 116–121, IEEE Computer Soc.共2000兲.
21. A. Piva, M. Barni, F. Bartolini, and V. Cappellini, “Threshold selec-tion for correlaselec-tion based watermark detecselec-tion,” in Proc. COST254
Workshop on Intelligent Communications, pp. 67–72共1998兲.
22. J. Dittmann, F. Neck, A. Steinmetz, and R. Steinmetz, “Interactive watermarking environments,” in IEEE Conf. on Multimedia
Comput-ing and Systems, pp. 286–294共1998兲.
