数字水印英文资料

Tamkang Journal of Science and Engineering, Vol. 4, No. 3, pp. 201-208 (2001) 201 Digital Image Multiresolution Watermark Based onHuman Visual System Using Error Correcting CodeChing-Tang Hsieh and Yeh-Kuang WuDepartment of Electrical EngineeringTamkang UniversityTamsui, Taipei, Taiwan 251, R.O.C.E-mail: hsieh@.twAbstractDigital watermarking has been proposed as a way to claimprotection. In this paper, we try to use the error-correcting code skill,the multiresolution of wavelet transformation, and the human visualsystem to improve the traditional watermarking. Error-correctingbased watermarking will have the property that corrects errors of theextracted watermark automatically. The HVS (Human Visual System)model will be presented to investigate perceptually characteristic partsof human (e.g. intensity or hue) and how to insert the watermark intothe original data, where the perceptual sensitivity of human isrelatively low. A multiresolution watermarking based on the wavelettransformation is selected in each frequency band of the DiscreteWavelet Transform (DWT) domain and therefore it can resist thedestruction of image processing. In our experiments, the results showthat the robustness of a watermark with ECC (Error Correct Coding) ismuch better than the traditional one without ECC.Key Words: DWT, Watermark, HVS, Error-correcting Code, BCH1. IntroductionThe rapid development of Internet introduces a new set of challenging problems regarding security. One of the most significant problems is to prevent unauthorized copying of digital production from distribution. Digital watermarking has provided a powerful way to claim intellectual protection [3,6,7,10,11,23-25]. Watermark must have two most important properties: transparency and robustness.y Transparency refers to the perceptual quality ofthe watermarked data. The watermark should beinvisible over all types.y The digital watermark is still presented in theimage after distortion and the watermark detectorcan detect it. Ideally, the amount of imagedistortion necessary to remove the watermarkshould degrade the desired image quality to thepoint of becoming commercially valueless. It iscalled the robustness of digital watermark toimage processing. The common distortion ofsignal processing includes lossy compression (inparticular JPEG), resampling, requantization,image enhancement, cropping, etc.A key point of the watermarking technique is the trade-off between the transparency and the robustness. We must determine where to insert watermark and how to enhance the robustness of the watermark. In this paper a wavelet based watermarking method using the HVS [11,24] is presented. In [14], Lee adopted the RS code (Reed-Solomon code) to generate ECC codewords, and regarded the codeword’s parities as watermark. They used watermark to recover the damaged image. But from the viewpoint of intellectual protection, the watermark is more important than the image, while the image quality is maintained. They did not handle the watermark.We proposed an idea for enhancing the robustness of extracted watermarks. Watermark can be treated as a transmitted signal, while the destruction from attackers is regarded as a noisy distortion in channel. According to the viewpoint mentioned above, we provide an idea using ECC to202 Ching-Tang Hsieh and Yeh-Kuang Wudetect and correct the error part of the extracted watermarks. The organization of this paper is as follows. Section 2 presents the schemes of this paper. Section 3 describes the watermark insertion and Section 4 presents the watermark detection. Experimental results and discussions are given in Section 5 and conclusions are drawn in Section 6.2. Method2.1 Human Visual System ModelThe experiment result indicates that human vision reacts with different sensitivity to each frequency band [4,13], which is divided into several narrow band channels. This property of human vision is represented by contrast sensitivity function: Modulation Transfer Function (MTF). Mannos and Sakrison’s work [18] may be the first major breakthrough in image coding incorporating the Human Visual System (HVS). They proposed the MTF to model the HVS.d cf cf b a f MTF )exp()()(−+=(1) where f is the radial frequency incycles/degree (CPD) of the visual angle subtended and a, b, c, and d are constants. In this paper, we will follow Mannos and Sakrison’s transfer function:1.1))114.0(exp()114.0192.0(6.2)(f f f MTF −+= (2) The transformation from spatial frequency domain of MTF to normalized frequency domain is shown in Figure 1.Because a human observer evaluates the quality of image, the properties of HVS will use as the weight estimator for enhancing the transparency effect of watermark [11].2.2 Error-Correction CodingError-correction coding theory was proposed in 1948. It was used for resisting the corrosion of noise in data communications [15,19]. We regard the inserted watermark as transmitted information and encode it by ECC before inserting. In this paper, we compare the result of extracted watermark with ECC and without ECC and we use two kinds of ECC for our experiments: the repetition coding and the BCH (Bose, Chaudhuri, and Hocquenghem) coding [2,5,15,19,21]. 2.2.1 Repetition CodingThe rule of repetition encoding is repeating each original signal of a watermark N times in a block section, named block section (N, 1). In the decoding process, we use the majority elements of the block section to reconstruct the original signal. For example, we set N=5 in the binary signal and the (00000) represents 0, the (11111) represents 1. In the decoding process, the reconstructed signal is ‘0’ if the number of ‘0’ is more than 3 in a block section; otherwise it is ‘1’.2.2.2 BCH CodingThe Bose-Chaudhuri-Hocquenghem (BCH) codes are a class of most extensively studied random-error- correcting cyclic codes [1,5]. The cyclic structure of BCH codes is the standard algebraic decoding method consisting of three major step [5,15]:(1) Calculate the syndrome valuest i S i 2,...,2,1,= form the received-word polynomial )(x r ;(2) Determine the error-locationpolynomial )(x σ from the syndrome values of the received word; and(3) Solve for the roots of )(x σ, which arethe error locators.In our experiment, we use a 3-error-correcting binary BCH for coding, whose block-length is 15-bits.3. Watermark InsertionThe block diagram of our digital watermarking method is shown in Figure 2.Function Normalized frequencyM a g n i t u d eFigure 1. Normalized Modulation TransferFunction.Digital Image Multiresolution Watermark Based on Human Visual System Using Error Correcting Code 203Figure 2. Digital watermarking methodFirst, we use a pseudo-random sequence as watermark W . W e decompose the original image by DWT in three levels to obtain the wavelet coefficient in each band as shown in Figure 3. In order to enhance the transparency of watermark, we multiply ith band by different weighting value [11,19,24]. The ith weighting value i w is the integral of a Modulation Transfer Function (MTF) over the frequency interval in each band:()∫∫=iiWB ss WB si df f H df w (3)where ()s f H and i WB are MTF and the bandwidth of ith band respectively.()()()ts s s sf rf q p f H −+=exp (4)⎟⎠⎞⎜⎝⎛=+i i i WB 21,211 (5)The procedure of insertion is described as follows: Step 1.We decompose the original image Y with athree-level DWT to obtain the wavelet coefficient y(m,n). The parameter m, n, represent the spatial location of each pixel in the decomposed image.Step 2.Calculate the weighting value i w of each bandaccording to equation (3), (4) and (5). The parameter p, q, r, s and t are constants in our experiment (p=2.6, q=0.192, r=s=0.114, t=1.1). Step 3.We encode the watermark W using ECCalgorithm to obtain W ′. Set a threshold T.Step 4.If the absolute magnitude of wavelet coefficientis larger than the threshold T (i.e. ≥),(n m y T)we insert the watermark into the wavelet coefficient y(m,n) according to the order shownin Figure 4.:(6) where ()n m y ,′and α are modified wavelet coefficient and parameter to control the level of the watermark respectively.Step 5.Finally, we take inverse DWT of the modifiedwavelet coefficient to obtain watermarked image Y ′.Figure 3. The decomposed Lena in three levelsFigure 4. The order of inserting the watermark4. Watermark DetectionWe can extract the watermark X essentially taking the inverse steps described in section 3. The extraction of watermark requires both original image Y and watermarked image Y ′. For watermark detection, a similarity measure ()W X sim , used in [6] is defined by eq.5:()XX W X W X sim ⋅⋅=, (5)where ‘⋅’ denotes the inner product of two vectors.5. System EvaluationWe use three test images in our experiments of which sizes are all 512 ×512 in pixels. The quantity of watermark actually inserted is 5000. The parameter α is set to 0.15 and a smaller value of 0.01α is set for the lowest frequency band LL 3. The threshold T is()()()W n m y w n m y n m y i ′⋅⋅⋅+=′,,,α204 Ching-Tang Hsieh and Yeh-Kuang Wuset to 4.We use two ECC algorithms and one mixed ECC algorithm in our experiments: I. Repetition coding:T he encoded watermark is 4 and 8 multiple of the original watermark respectively. II. 3-error-correcting binary BCH coding.T he encoded watermark is 3 multiple of the original watermark.III. 3-error-correcting binary BCH coding togetherwith repetition 4 times coding (Mixed-BCH coding).T he encoded watermark is 12 multiple of the original watermark. 5.1 Uniqueness of the WatermarkFigure 5 shows the detection response Sim of the watermark detector to 1000 random watermarks, among which there is only one matches the watermark we actually insert and the detection response is 70.71.5.2 PSNR of the Watermarked ImagesWe evaluate the PSNRs of the three watermarked test images with and without various ECC algorithms. Figure 6 (a), (b) and (c) show the three test images and Figure 6 (d), (d) and (f) show the watermarked versions of Figure 6 (a), (b) and (c). Table 1 tabulates the PSNRs of the three watermarked images with different ECC algorithms. Even if the three images are watermarked in different ECC algorithms, the PSNRs are very high (not smaller than 34) and the image qualities are still very well (Figure 6 (d), (e) and (f)).When the watermarked image are compressed into JPEG images (with quality 100), the detection responses Sim are tabulated in Table 2. We can distinguish the owner according to the enough high detection responses. Especially when we use the BCH+repeat coding, the responses of the watermarked images almost reach to the peak value (71.71). 5.3 JPEG CompressionWe demonstrate the JPEG compression of the watermarked images with different JPEG qualities. Figure 7 shows the watermarked Lena and the JPEG compression of the watermarked Lena. Table 3 tabulates the detection responses of the watermarked images under different JPEG qualities. No matter which ECC algorithm we adopt, the detection responses are much better than the responses without using ECC algorithm. Figure 8 shows the results graphically for easily distinguishing the differences between various ECC algorithms.Even though when the JPEG qualities are smaller than 20, the responses of the watermarked image can provide a certification of the owner. However when the JPEG quality downs to 40 or more below, the watermarked images are strictly destructed and commercially valueless.(a)Original Lena (b)Original Cat (c)OriginalFlower(d)Watermarked Lena (e)Watermarked Cat (f)Watermarked FlowerD e t e c t i o n r e s p o n s e S i mFigure 5. Response of the watermark detector to 1000 random watermarks1000 random watermarkDigital Image Multiresolution Watermark Based on Human Visual System Using Error Correcting Code 205 Figure 6. Original images (a)Lena (b)Cat (c)Flower and watermarked images using BCH coding (d)Watermarked Lena(e)Watermarked Cat and (f)Watermarked Flower206 Ching-Tang Hsieh and Yeh-Kuang WuTable 1. The PSNR of watermarked images with variousECC algorithmsTable 2. Detection responses of the three test images with various ECC algorithm. (JPEG quality=100)Table 3. Detection responses of the watermarked Lena with various ECC algorithm under different JPEG qualitiesRepetition code Images Without ecc N=4N=8BCHcodeBCH code +repetition code (N=4) Watermark length k 4k 8k 3k 12k PSNR of watermarkedLena 52.25 39.9938.94 40.30 38.80 PSNR of watermarkedCat 51.69 43.4042.56 43.70 42.42 PSNR of watermarked Flower46.43 36.1234.99 36.5334.83Repetition code ImagesWithout ecc N=4N=8 BCH code BCH code +repetition code Lena 11.74 55.4463.07 50.91 70.43 Cat 12.42 56.7963.87 51.76 70.71 Flower 14.09 57.0863.98 52.27 70.71 Repetition codeJPEG QualityCompress ratioWithout eccN =4 N =8BCH codeBCH code + repetition code(N=4)100 1.69 11.74 55.44 63.07 50.91 70.43 90 4.60 10.66 51.28 56.06 46.81 55.92 80 7.08 9.11 44.35 49.04 41.55 45.20 70 9.038.6638.6142.7738.2737.56 60 10.92 7.38 35.02 38.13 32.84 31.42 50 12.48 6.90 33.15 34.87 29.70 25.68 40 14.56 6.31 29.39 31.20 26.36 23.65 30 16.38 4.92 23.96 24.58 22.17 17.48 20 20.15 3.51 19.32 18.41 15.92 10.89 10 29.11 1.41 9.64 9.53 6.76 4.67 043.67- 0.65- 0.25- 0.140.25D e t e c t i o n r e s p o n s e S i mFigure 8. Detection response of the watermarked images with various ECC algorithm under different JPEG qualities JPEG qualityDigital Image Multiresolution Watermark Based on Human Visual System Using Error Correcting Code 207(a)Watermarked Lena (b)JPEG compression of (a)Figure 7. (a) The watermarked Lena and, (b) the JPEG compression of the watermarked LenaTable 4. Detection responses with lowpass filtering onceand twiceTable 5. The PSNR of watermarked images comparingwith various ECC algorithmsTable 6. Detection responses with some geometricimage processing5.4 Geometric AttacksRobustness against geometric manipulation, such as filtering, scaling, or stirmark, is very important because these manipulations are very common and usually do not degrade too much the quality of the image. And here we present the robustness with some quite general kinds of geometric manipulation.Table 4 tabulates the results with a 3x3 averaging filtering once and twice. The Table 4 shows that the order of the nice quality is: repetition coding (4,1), BCH, repetition coding (8,1), and Mixed-BCH coding. It shows that too much extra coding will induce the most part of a watermark inserted into the wavelet coefficient in higher band and easily destroyed by an average filter. Consequently, the length of ECC code should be as short as possible.The results of comparing to the previous method with ECC are shown in Table 5. In Niu’s method [20], a decomposed watermark is added to the subbands of the DWT decomposition of the host image including the highest frequency subband. Each bitplane of the gray-scale watermark is embedded directly with the hamming code. Although the method can be hidden much information than ours, we prefer to maintain the high image quality according to the imperceptibility, the basic requirement of watermarking scheme.Table 6 shows the results with some general attacks. StirMark [1,22] is a generic tool developed for simple robustness testing of image marking algorithms and it simulates a re-sampling process. These simple transformations are more likely to be used and include: horizontal flip, scaling (with the factor of 0.5), 3x3 median filter, mosaic and StirMark 1.0.6. ConclusionIn this paper, we propose an idea to enhance the robustness of the watermarkusing the ECC algorithm. According to thesensitivity of human, HVS keeps awatermarked image quality from thedistortion induced by watermark-insertionmethods. With the characteristic of ECC, thedetection response of the processed image isrelatively higher than the traditional one without using the ECC algorithm. In this experiment, we demonstrate the performance Repetitioncode Times of lowpass filtering Without ecc N=4N=8BCH codeBCH code + repetition code(N=4) 1 33.40 34.9923.08 33.06 12.78 2 23.02 23.0513.52 20.05 4.04==Repetition code N=4 BCH code Median Filter35.8030.27Scaling 32.64 32.35 StirMark 26.66 24.82 Mosaic 35.50 34.42 Horizontal flip 48.25 43.68208 Ching-Tang Hsieh and Yeh-Kuang Wuof the proposed method by computer simulation. Further research will focus on the development of robust watermarking methods with more powerful error-correcting capability algorithm.References[1]Barni, M., Bartolini, F. and Piva, A.,“Improved wavelet-based watermarkingthrough pixel- wise masking,” ImageProcessing, IEEE Trans., Vol. 10, pp.783-791 (2001).[2]Blahut, R. E., Theory and Practice ofError Control Codes, Reading,Addison-Wesley (1983).[3]Bors, A. G. and Pitas, I., “Imagewatermarking using DCT domainconstraints,” Proc. of Int. Conf. onImage Processing, Vol. 3, pp. 231-234(1996).[4]Chitprasert, B., Rao, K. R., “Humanvisual weighted progressive imagetransmission,” Communications, IEEETrans., Vol. 38, pp. 1040-1044 (1990). [5]Clark, C. C. and Cain, J. B., ErrorCorrecting Coding for Digital Communication, Plenum (1981).[6]Cox, I. J., Kilian, J., Leighton, T. andShamoon, T., “Secure spread spectrumwatermarking for multimedia,” ImageProcessing, IEEE Trans., Vol. 6, pp.1673- 1687 (1997).[7]Cox, I. J., Kilian, J., Leighton, T. andShamoon, T., “Secure spread spectrumfor multimedia,” Image Processing 97,IEEE Trans., Vol. I, pp. 548-551(1997). [8]Haykin, S., Communication Systems 3/e,John Wiley & Sons, U. S. A. (1994).[9]Hsieh, C. T., Lu, Y. L., Luo, C. P. and Kuo,F. J., “A study of enhancing therobustness of watermark,” Proc. IEEE Int.Sym. on Multimedia Software Engineering,Taiwan, pp. 325-327 (2000).[10]Hsu, C. T. and Wu, J. L., ”Multiresolutionwatermarking for digital images,”Circuits and Systems II: Analog andDigital Signal Processing, IEEE Trans.,Vol. 45, pp. 1097-1101 (1998).[11]Kim, Y. S., Kwon, O. H. and Park, R. H.,“Wavelet based watermarking method fordigital images using the human visualsystem,” Proc. IEEE Int. Sym. on ISCAS'99,Orlando, FL, U. S. A., Vol. 4, pp.80-83 (1999). [12]Kuo, C. J. and Chen, M. S., “A newsignal encryption technique and its attackstudy, Proc. Int. Conf. on SecurityTechnology,” Taipei, Taiwan, pp.149-153 (1991).[13]Lee, I., Kim, J., Kim, Y., Kim, S., Park, G.and Park, K. T., “Wavelet transformimage coding using human visualsystem,” Proc.IEEE Circuits andSystems, Taiwan, pp. 619-623 (1994).[14]Lee, J, Won, C. S., “Authentication andcorrection of digital watermarkingimages,” Electronics Letters, Vol. 35, pp.886-887 (1999).[15]Lin, S. and Costello, D. J., Error ControlCoding, Prentice Hall (1983).[16]Lu, Y. L., “Static image security withlow human visual respond characteristic& multi-frequency wavelet watermark,”Master Thesis, University of Tamkang,Taiwan, R. O. C. (2001).[17]Luo, C. P.,“A study of the robustness ofwatermark,” Master Thesis, University ofTamkang, Taiwan, R. O. C. (2001).[18]Mannos, J. L. and Sakrison, D. J., “Theeffect of a visual fidelity criterion onthe encoding of images,” IEEEInformation Theory, Vol. IT-20, pp.525-536 (1974).[19]Michelson, A. M. and Levesque, A. H.,Error-Control Techniques for DigitalCommunication, Wiley, U. S. A. (1985).[20]Niu, X. M., Lu, Z. M. and Sun, S. H.,“Digital image watermarking based onmultiresolution decomposition,”Electronics Letters, Vol. 36, pp.1108-1110 (2000).[21]Peterson, W. W. and Weldom, E. J.,Error- Correcting Codes, M.I.T. Press,Cambridge, M.A., U. S. A. (1972).[22]Petitcolas, F. A. P., Anderson, R. J. andKuhn, M. G., “Attacks on copyrightmarking systems,” Proc. Int. onInformation Hiding, Portland, Oregon, U.S. A., pp. 219-239 (1998).[23]Piva, A., Barni, M., Bartolini, F. andCappellini, V., “DCT-based watermarkrecovering without resorting to theuncorrupted original image,” Proc. Int.Conf. on Image Processing, Vol. 1, pp.520-523 (1997).[24]Wei, Z. H., Qin, P. and Fu, Y. Q.,“Perceptual digital watermark of imagesusing wavelet transform,” ConsumerDigital Image Multiresolution Watermark Based on Human Visual System Using Error Correcting Code 209 Electronics, IEEE Trans., Vol. 44, pp.1267-1272 (1998).[25]Xia, X. G., Boncelet, C. G. and Arce, G.R., “A multiresolution watermark fordigital images,” Proc. Int. Conf. onImage Processing 97, Santa Barbara, CA,U. S. A. Vol. I, pp. 548-551, (1997).Manuscript Received: Jul. 15, 2001Revision Received: Aug. 15, 2001and Accepted: Aug. 20, 2001。