Kĩ thuật DWT in matlab
2013 10th International Mu lti-Conference on Systems, Signals & De vices (SSD) Hammamet, Tunisia, Marc h 18-21, 2013 SSD'13 1569695081 A Robust Multiple Watermarking Scheme Based on the DWT Ouazzane Hana, Mahersia Hela, Hamrouni Kamel Université de Tunis El Manar, Ecole Nationale d'Ingénieurs de Tunis LR-SITI: Signal Image et Technologies de l'Information, Tunis, Tunisia Abstract— In this paper we make contributions to a non-blind multiple watermarking scheme that proceeds b y embedding a binary image in the discrete wavelet transform b ands of a gray scale image. Unlike the common wavelet based watermarking techniques, the proposed scheme lies essentially o n marking the approximation and diagonal bands of the dis crete wavelet transform (DWT) of the cover image achievi ng a better compromise between fidelity and robustness. Ex periments show that our contributions provide the multiple watermarking scheme with robustness to a wide variety of attack s. Index Terms— Digitalwatermarking, discret e wavelet transform, non-blind image watermarking. I. INTRODUCTION The development of communication networ ks and the trivialization of image processing tools have gi ven rise to content security problems underscoring the nee d to secure digital images from illegal modification, protect th eir economic interest and ensure intellectual property. Digital image watermarking is an attractive alt ernative that matches these necessities.This technique consists in embedding a permanent watermark in a cover ima ge in such a way that the watermarked image remains ac cessible to everyone and the embedded watermark can be d ecoded after the watermarked image have undergone seve ral attacks. Besides, potential attacks can be no-malicious like compression and image enhancement techniques or malicious like rewatermarking and cropping [1] [2]. The embedd ed mark can be visible or invisible. Digital watermarking has many applications a ccording to the type of the watermark and the used technique. In general, visible watermarking is used to reveal ownershi p, invisible robust watermarking is used for copyright pro tection and organization of digital contents in archiving sy stems, and, invisible fragile watermarking is used for tamperin g detection. Image watermarking requires usually three rele vant criteria [3]: - Fidelity: the watermarking process should not distort the original image to ensure its commercial value. - Robustness: the inserted mark should be detec table if the cover image has undergone some potential attacks. It should be, however, difficult and complex to be detected by unauthorized people. Fragile watermarks shoul d be altered in an irreversible way if the cover image has bee n modified. - Capacity: it describes the necessary amount of data to be inserted in the cover image. Watermarking sche mes should have high capacity. Every watermarking scheme includes an encod er, process that embeds the watermark in the cover image, and a decoder, process that detects or extracts the watermark. W atermarking schemes can be distinguished according to the en coding and decoding domain. Effectively, images can be repres ented either in the spatial domain i.e. the image pixel do main, or a transformed domain such as discrete cosine transf orm domain or discrete wavelet transform domain. Watermark embedding in the spatial domain is performed by modifying the cover image pixels values. Watermark embedding in a t ransformed domain is performed by modifying the image coe fficients in this selfsame domain. Watermarking scheme s can be distinguished according to the watermark embeddin g approach: - LSB substitution Approach: embeds the wat ermark by substitution of some specific least significant bi ts (LSB) of the cover image pixels, like the schemes [4], [5] and [6]. - Additive insertion: adds the watermark to so me image components, like the schemes [7], [8] and [9]. - Statistical approach: this approach is known as Patchwork [10], it performs by pseudo randomly choosing pixels from the cover image and modifying their luminosity. - Visual approach: Texture block watermarking i s a method that lies on the visual approach. It uses rando m texture patterns in the cover image. It performs by producing identical textured regions by copying a rando mly chosen pattern [10]. - Quantization based watermarking: This appr oach uses quantization to embed the watermark in the co ver image. For example, the scheme proposed in [11] pe rforms by quantizing coefficients relative to some special i mage edges to embed the binary watermark bits. 978-1-4673-6457-7/13/$31.00 ©2013 IEEE 1 Watermarking schemes can be also classified a ccording to the decoder type. There are three decoding modes: - Non-blind decoding: requires at least the original image. - Semi-blind decoding: uses only the original watermark. - Blind decoding: extracts the watermark from t he possibly distorted image using neither the original im age nor the original watermark. In this paper, we present a new watermarking sc heme based on the DWT. This transform is commonly use d in digital watermarking because of its advantages. Keyv anpour and Merrikh-Bayat propose in [11] a blind watermar king scheme that embeds the watermark in the HL and LH sub-bands resulting from a multilevel DWT using the q uantization approach. Tao and Eskicioglu propose in [12] a non-blind multiple watermarking scheme based on the DW T. First, they apply the first or second level DWT to the cover image. The level choice depends on the watermark size that m ust be equal to each sub-band thumbnail size. According to the additive approach, they embed four copies of the binary wa termark into the LL, HL, LH and HH sub-bands. They apply the IDWT to get back to the spatial domain and obtain the w atermarked image. The decoding process consists in applyin g the DWT and extracting the four embedded watermark copi es. The four extracted watermarks are, afterwards, compared t o the original watermark to check the watermark presence in t he attacked image. For objective examination, they calculate t he similarity ratio (SR) between each extracted watermark and the original one and admit that the highest SR value helps to identify the most resistant sub-band for a given attack. Our c ontributions consist in embedding only two copies of the water mark in the High and Low frequency sub-bands. In fact, th e extraction results in [12] show that the highest SR values are always found at the LL or HH sub-bands according to the attack type. The paper continuous as follows: In section 2 we present a brief introduction to the two dimensional DW T and we describe the encoding and decoding processes. Section 3 is dedicated for experimental evaluation. In this section we present our test platform and the results of t he method simulation and we compare the new scheme to so me schemes based on the DWT. Finally, in section 4, we give our observations regarding the obtained scheme simul ation results and our perspectives. II. PROPOSED WATERMARKING M ETHOD Every two-dimensional DWT decomposition level produces four representations of an image: an ap proximation image (LL) and three detail ones (LH, HL and HH). The approximation image represents the image low fr equencies, it has the largest coefficient magnitudes at each lev el and, thus, contains the most significant information of the image. To obtain the next level decomposition, the two dimen sional DWT is applied to the LL sub-band. The detail images are called the vertical (LH), the horizontal (HL) and the diagon al (HH) sub- bands, they represent the mid and high frequenc y sub-bands and contain information about edges and textur e patterns. Figure 1 shows a two level decomposition. Fig. 1. Two level DWT decompositi on In the following, we describe the embeddin g and the extraction process. A. Watermark embedding process The cover image (I) is a gray scale image. We s uppose that the cover image size is: N*N, then the binary water mark image (W) size must be ; n is the decomposition le vel during the embedding process. 1. Decompose I using the n- level two dimensional DWT. 2. Inserting the watermark in t he LLn and HHn sub-bands by modifying their coefficients as follows : ( ) ( ) denotes the sub-band LLn or HHn. is the watermarked LLn or HHn image repre sentation. denotes the scaling factor corresponding to each sub- band. Effectively, we don’t use the same sca ling factor for the LLn and HHn sub-bands since the coefficient sizes are not of the same magnitude order. 3. Apply the n-level IDWT to obtain the watermarked image Î in the spatial domain. B. Watermark extraction process Let I’ be the possibly corrupted image. 1. Decompose I’ using the n-level two dimensional DWT. 2. Extracting the watermark from the LLn and HHn sub- bands as follows : ( ) ( ( ) ( )) ⁄ is the extracted watermark from the LLn o r HHn sub- bands. 3. Conver t to a bi nary image a pplying a si mple thresholding : 2 ( ) { ( ) III. EXPERIMENTAL EVALUATI ON In this part , we test the p roposed wate rmarking sch eme on the 512*512 gra y scale Goldhill test image and we use three binary watermar ks (see fig. 2). Effectively the G oldhill test image is used in Tao and Eskiciog lu’s paper, thus, i t will be First level decomposition Second level decompositio LL: SR = 1.000HH: SR = 1.00LL: SR = 1.000HH: SR = 1.00 PSNR(d b) useful to compare the proposed scheme with Tao and Eskicioglu’s method. The tests will involve wate rmarking of the LL and HH sub-bands for first and second level DWT decomposition. To evaluate the proposed scheme fidelity, we measure the visual quality of the watermarked image using the Peak Signal to Noise Ratio (PSNR). (b) Watermar k used to test robustnes s against attacks. (a) Goldhill cover test image (c) Watermark used for rewatermarking attack. (d) Watermark use d to compare the prop osed scheme with Tao and Eskicioglu’s sche me. Where, the RMSE is the square root of mean squ ared error (MSE) between the original image and the distorted one. √ ∑ ∑ [ ( ) ( )] Qualitative evaluation of the watermark presen ce can be done by comparing the two extracted watermarks with the original one. Quantitativeevaluationis performed by calculating the similarity ratio (SR) between each extracted watermark and the original one. The SR value lies between 0 and 1. Where, S is the number of matching pixels bet ween the original and extracted watermarks, and D is the n umber of different pixels between the same images. Figure 3 presents the PSNR values of twel ve first level watermarked images (Goldhil l, Lena, Peppers, Couple, Cameraman, Boat, F16, Barbara, Mandrill, Prin ter test, Zelda Fig. 2. Test platform 50 40 30 20 10 0 Fig.3 . The PSNRs of twelve watermarked gray sc ale images each of size 512*512. and Pirate). The figure shows that all the PSNR values are greater than 40 db. Figure 4 presents the results of embedding t he “WMK” (a) Watermarked image at first level decomposition, PSNR = 42.724 db. (b) Watermarked image at first level decomposition, PSNR = 42 .723 db. binary logo into the Goldhill test image (see fig. 2). Embedding PSNRs indicated in Tao and Eskicioglu’s paper : Eskicioglu’s method. Eskicioglu’s method. Fig. 4. Watermarking results. Fig. 5. Watermark extraction results. 3 Figure 5 provides the extraction results witho ut applying any attack on the watermarked image. To evaluate the scheme robustness, we ha ve applied different attacks to the watermarked Goldhill ima ge. For each attacked image, we have extracted the two embedded watermarks and calculated the SRs. Figures 6 and 7 provide the extracted watermarks from the LL and HH sub- bands and their appropriate SR after each attack. These results sug gest that: - The LL sub-bands aremost resistant to lossy compression, filtering, geometrical deform ations and noise addition. - The HH sub-bands are robust to nonlinear d eformations of the gray scale. - Both sub-bands are resistant to rewatermarking. - Robustness is enhanced for second level dec omposition. In particular, the visual quality of LL extracted watermarks and their SR values have be en visibly increased in fig. 7 for the lossy compressio n, low-pass filtering, sharpening and noise addition. Comparison with previous methods In this part, we compare the experimental res ults of the proposed method with Tao and Eskicioglu’s method and Yuan’s method [13]. These two methods are ba sed on the multiple watermarking approach in the DWT do main. Results the “BC” binary logo, used in Tao and Eskicioglu’s pap er, in the Goldhill image gives PSNR values exceeding slig htly - First level decomposition: PSNR = 42.724 db with proposed scheme vs. PSNR = 42.400 db with Tao and - Second level decomposition PSNR = 42.701 db with proposed scheme vs. PSNR = 42.230 db with Tao and are shown in figures 8, 9, 10, 11, 12 and 13, they a re based on applying the same attacks on the Goldhill t est image watermarked with the same binary logo for each co mparison. Figures 8, 10 and 12 provide the SR values after watermark extraction from the LL sub-band. They show that the SRs of the proposed method exceed the SRs of both previo us methods. In particular, the robustness of the LL sub-band has improved significantly for the gray scale deformation attac ks such as histogram equalization. Figures 9, 11 and 13 provide the SR val ues after watermark extraction from the HH sub-band. The SR values reveal also that the HH sub-band robustness is enhanced with the proposed scheme. IV. CONCLUSION In this paper, we have made a contribution to a multiple non-blind watermarking scheme based on the DWT. The proposed scheme consists in applying the DWT t o the gray scale cover image and modifying the LL and H H sub-band coefficients in order to insert the binary watermar k according to an additive approach. Experimental results indicate that modification of the LL and HH sub-bands results in good fidelity and robustness against a large range of attacks. Watermark embe dding with second level decomposition results in better r obustness. Objective evaluation shows that the propose d method outperforms Tao and Eskicioglu’s scheme in term s of fidelity and robustness. The proposed watermarking method can b e further improved by automating the selection of the optimal thresholding parameter and appropriate scaling fac tor for each band. JPEG Compression Q=25 JPEG Compression Q=50 JPEG Compression Q=75 Gaussian filtring (3 * 3) LL: SR = 0.813 HH: SR = 0.477 LL: SR = 0.899 HH: SR = 0.478 LL: SR = 0.958 HH: SR = 0.476 LL: SR = 0.879 HH: SR = 0.476 Gaussian filtring (5 * 5) Sharpening Histogram equalization Intensity Adjustment ([0 0.8][0 1]) LL: SR = 0.773 HH: SR = 0.477 LL: SR = 0.936 HH: SR = 0.916 LL: SR = 0.682 HH: SR = 0.885 LL: SR = 0.855 HH: SR = 0.897 Gamma correction (1.5) Pixelate Gaussian noise ([0 0.001]) Rescaling (512 -> 256 -> 512) LL: SR = 1.000 HH: SR = 1.000 LL: SR = 0.800 HH: SR = 0.475 LL: SR = 0.782 HH: SR = 0.653 LL: SR = 0.903 HH: SR = 0.476 Cropping Rewatermarking LL: SR = 0.863 HH: SR = 0.920LL: SR = 0.880 HH: SR = 0.880 Fig. 6. Extracting results for first decomposition level. 4 JPEG Compression Q=25 JPEG Compression Q=50 JPEG Compression Q=75 Gaussian filt ring (3 * 3) LL: SR = 1.000 HH: SR = 1.000 LL: SR = 0.985 HH: SR = 0.607 LL: SR = 0.991 HH: SR = 0.828 LL: SR = 0.971 HH: SR = 0.634 Gaussian filtring (5 * 5) Sharpening Histogram equalization Intensity Adjustment ([0 0.8][0 1]) LL: SR = 0.893 HH: SR = 0.442 LL: SR = 0.992 HH: SR = 0.931 LL: SR = 0.691 HH: SR = 0.884 LL: SR = 0.853 HH: SR = 0.895 Gamma correction (1.5) Pixelate Gaussian noise ([0 0.001]) Rescaling (512 -> 256 -> 512) LL: SR = 1.000 HH: SR = 1.000 LL: SR = 0.860 HH: SR = 0.512 LL: SR = 0.928 HH: SR = 0.774 LL: SR = 0.981 HH: SR = 0.636 Cropping Rewatermarking LL: SR = 0.877 HH: SR = 0.926LL: SR = 0.885 HH: SR = 0.885 Fig. 7. Extracting results for second decomposition level. 0,9 0,7 0,5 0,3 0,1 Tao and Eskicioglu's method Proposed method a b c d e f g h i j k 0,9 0,7 0,5 0,3 0,1 Tao and Eskicioglu's method Proposed method a b c d e f Fig. 8. LL sub-band robustness comparison betwe en Tao’s method and the proposed method for first level decomposition. (a) JPEG compression (Q=25), (b) JPEG compression (Q=50), (c) JPEG compression (Q=75), (d) Gaussian filtering (3*3), (e) Sharpening, (f) rescaling (512->256- >512), (g) Gaussian noise ([0 0.001]), (h) Histog ram equalization, (i) Intensity adjustment ([0 0.8][0 1]), (j) Gamma correction (1.5), (k) Rewatermarking. Fig. 9. HH sub-band robustness comparison bet ween Tao’s method and the proposed method for first level decompo sition. (a) Histogram equalization, (b) Intensity adjustment ([0 0,8], [0 1]), (c) Gamma correction (1,5), (d) Sharpening, (e) Gaussian noise ([0 0.001]), (f) Rewatermarking. 5 0,9 0,7 0,5 0,3 0,1 Tao and Eskiciolu's method Proposed method a b c d e f g h i j k 0,9 0,7 0,5 0,3 0,1 Yuan's method Proposed method a b c d e f Fig. 10. LL sub-band robustness comparison between Tao’s method and the proposed method for second level deco mposition. (a) JPEG compression (Q=25), (b) JPEG compressio n (Q=50), (c) JPEG compression (Q=75), (d) Gaussian filtering (3 *3), (e) Sharpening, (f) rescaling (512->256->512), (g) Gaussian noise ([0 0.001]), (h) Histogram equalization, (i) Intensity adjustment ([0 0.8][0 1]), (j) Gamma correction (1.5), (k) Rewatermarking. 0,9 0,7 0,5 Fig. 13. HH sub-band robustness comparison b etween Yuan’s method and the proposed method for first level deco mposition. (a) JPEG compression (Q=75), (b) Histogram equalization , (c) Intensity adjustment ([0 0,8], [0 1]), (d) Gamma correction (1,5), ( e) Gaussian noise ([0 0.001]), (f) Cropping. REFERENCES [1] V. M. Potdar, S. Han and E. Chang, “A su rvey of digital image watermarking techniques,” Industrial Informatics, 3rd IEEE International Conference on, pp. 709- 716, 2005. [2] S. P. Mohanty, “Digital watermarking : a t utorial review”, unpublished. [3] E. Ganic and A. M. Eskicioglu, “Robust D WT-SVD domain 0,3 0,1 Tao and Eskicioglu's method Proposed method a b c d e f image watermarking : embedding data in al l frequencies,” Proceedings of the 2004 workshop on M ultimedia and Security, pp. 166-174, 2004. [4] P-Y Chen and H-J Lin, “A DWT based app roach for image steganography,” International Journal of App lied Science and Fig. 11. HH sub-band robustness comparison between Tao’s method and the proposed method for second level decomp osition. (a) Histogram equalization, (b) Intensity adjustment ([0 0,8] , [0 1]), (c) Gamma correction (1,5), (d) Sharpening, (e) Gaussian noise ([0 0.001]), (f) Rewatermarking. 0,9 0,7 Engineering 4, pp. 275-290, 2006. [5] A. Bamatraf, R. Ibrahim and M. N. M. Salleh , “A new digital watermarking algorithm using combination of least significant bit (LSB) and inverse bit,” Journal of Computi ng, vol. 3,2011. [6] R. G. Van Schyndel, A. Z. Tirkel and C. F. Osborne, “A digital watermark,” Proceedings. ICIP-94., IEEE International Conference, vol. 2, pp. 86-90, 1994. [7] E. T. Lin and E. J. Delp, “Spatial synchr onization using watermark key structure,” Security, Stega nography and Watermarking of Multimedia Contents, pp. 5 36-547, 2004. [8] P. Bas, B. Roue and J-M Chassery, “Tato uage d’images 0,5 0,3 Yuan's method Proposed method . Robust Multiple Watermarking Scheme Based on the DWT Ouazzane Hana, Mahersia Hela, Hamrouni Kamel Université de Tunis El Manar, Ecole Nationale d'Ingénieurs. d’insertion optimal,” Coresa03, 2003. [9] S. Rastegar, F. Namazi, K. Yaghmaie and A. Aliabadian, “Hybrid watermarking algorithm based on singular value 0,1 a