2013 a robust multiple watermarking scheme based on the DWT
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 60 61 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 by embedding a binary image in the discrete wavelet transform bands of a gray scale image. Unlike the common wavelet based watermarking techniques, the proposed scheme lies essentially on marking the approximation and diagonal bands of the discrete wavelet transform (DWT) of the cover image achieving a better compromise between fidelity and robustness. Experiments show that our contributions provide the multiple watermarking scheme with robustness to a wide variety of attacks. Index Terms— Digital watermarking, discrete wavelet transform, non-blind image watermarking. I. INTRODUCTION The development of communication networks and the trivialization of image processing tools have given rise to content security problems underscoring the need to secure digital images from illegal modification, protect their economic interest and ensure intellectual property. Digital image watermarking is an attractive alternative that matches these necessities. This technique consists in embedding a permanent watermark in a cover image in such a way that the watermarked image remains accessible to everyone and the embedded watermark can be decoded after the watermarked image have undergone several attacks. Besides, potential attacks can be no-malicious like compression and image enhancement techniques or malicious like rewatermarking and cropping [1] [2]. The embedded mark can be visible or invisible. Digital watermarking has many applications according to the type of the watermark and the used technique. In general, visible watermarking is used to reveal ownership, invisible robust watermarking is used for copyright protection and organization of digital contents in archiving systems, and, invisible fragile watermarking is used for tampering detection. Image watermarking requires usually three relevant criteria [3]: - Fidelity: the watermarking process should not distort the original image to ensure its commercial value. - Robustness: the inserted mark should be detectable if the cover image has undergone some potential attacks. It should be, however, difficult and complex to be detected by unauthorized people. Fragile watermarks should be altered in an irreversible way if the cover image has been modified. - Capacity: it describes the necessary amount of data to be inserted in the cover image. Watermarking schemes should have high capacity. Every watermarking scheme includes an encoder, process that embeds the watermark in the cover image, and a decoder, process that detects or extracts the watermark. Watermarking schemes can be distinguished according to the encoding and decoding domain. Effectively, images can be represented either in the spatial domain i.e. the image pixel domain, or a transformed domain such as discrete cosine transform 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 transformed domain is performed by modifying the image coefficients in this selfsame domain. Watermarking schemes can be distinguished according to the watermark embedding approach: - LSB substitution Approach: embeds the watermark by substitution of some specific least significant bits (LSB) of the cover image pixels, like the schemes [4], [5] and [6]. - Additive insertion: adds the watermark to some 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 is a method that lies on the visual approach. It uses random texture patterns in the cover image. It performs by producing identical textured regions by copying a randomly chosen pattern [10]. - Quantization based watermarking: This approach uses quantization to embed the watermark in the cover image. For example, the scheme proposed in [11] performs by quantizing coefficients relative to some special image edges to embed the binary watermark bits. SSD'13 1569695081 1 2013 10th International Multi-Conference on Systems, Signals & Devices (SSD) Hammamet, Tunisia, March 18-21, 2013 978-1-4673-6457-7/13/$31.00 ©2013 IEEE 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 60 61 Watermarking schemes can be also classified according 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 the possibly distorted image using neither the original image nor the original watermark. In this paper, we present a new watermarking scheme based on the DWT. This transform is commonly used in digital watermarking because of its advantages. Keyvanpour and Merrikh-Bayat propose in [11] a blind watermarking scheme that embeds the watermark in the HL and LH sub-bands resulting from a multilevel DWT using the quantization approach. Tao and Eskicioglu propose in [12] a non-blind multiple watermarking scheme based on the DWT. First, they apply the first or second level DWT to the cover image. The level choice depends on the watermark size that must be equal to each sub-band thumbnail size. According to the additive approach, they embed four copies of the binary watermark into the LL, HL, LH and HH sub-bands. They apply the IDWT to get back to the spatial domain and obtain the watermarked image. The decoding process consists in applying the DWT and extracting the four embedded watermark copies. The four extracted watermarks are, afterwards, compared to the original watermark to check the watermark presence in the attacked image. For objective examination, they calculate the 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 contributions consist in embedding only two copies of the watermark in the High and Low frequency sub-bands. In fact, the 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 DWT 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 the method simulation and we compare the new scheme to some schemes based on the DWT. Finally, in section 4, we give our observations regarding the obtained scheme simulation results and our perspectives. II. PROPOSED WATERMARKING METHOD Every two-dimensional DWT decomposition level produces four representations of an image: an approximation image (LL) and three detail ones (LH, HL and HH). The approximation image represents the image low frequencies, it has the largest coefficient magnitudes at each level and, thus, contains the most significant information of the image. To obtain the next level decomposition, the two dimensional DWT is applied to the LL sub-band. The detail images are called the vertical (LH), the horizontal (HL) and the diagonal (HH) sub- bands, they represent the mid and high frequency sub-bands and contain information about edges and texture patterns. Figure 1 shows a two level decomposition. Fig. 1. Two level DWT decomposition In the following, we describe the embedding and the extraction process. A. Watermark embedding process The cover image (I) is a gray scale image. We suppose that the cover image size is: N*N, then the binary watermark image (W) size must be ; n is the decomposition level during the embedding process. 1. Decompose I using the n-level two dimensional DWT. 2. Inserting the watermark in the LL n and HH n sub-bands by modifying their coefficients as follows : denotes the sub-band LL n or HH n . is the watermarked LL n or HH n image representation. denotes the scaling factor corresponding to each sub- band. Effectively, we don’t use the same scaling factor for the LL n and HH n 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 LL n and HH n sub- bands as follows : is the extracted watermark from the LL n or HH n sub- bands. 3. Convert to a binary image applying a simple thresholding : 2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 60 61 III. EXPERIMENTAL EVALUATION In this part, we test the proposed watermarking scheme on the 512*512 gray scale Goldhill test image and we use three binary watermarks (see fig. 2). Effectively the Goldhill test image is used in Tao and Eskicioglu’s paper, thus, it will be useful to compare the proposed scheme with Tao and Eskicioglu’s method. The tests will involve watermarking 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). Where, the RMSE is the square root of mean squared error (MSE) between the original image and the distorted one. Qualitative evaluation of the watermark presence can be done by comparing the two extracted watermarks with the original one. Quantitative evaluation is 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 between the original and extracted watermarks, and D is the number of different pixels between the same images. Figure 3 presents the PSNR values of twelve first level watermarked images (Goldhill, Lena, Peppers, Couple, Cameraman, Boat, F16, Barbara, Mandrill, Printer test, Zelda and Pirate). The figure shows that all the PSNR values are greater than 40 db. Figure 4 presents the results of embedding the “WMK” binary logo into the Goldhill test image (see fig. 2). Embedding the “BC” binary logo, used in Tao and Eskicioglu’s paper, in the Goldhill image gives PSNR values exceeding slightly PSNRs indicated in Tao and Eskicioglu’s paper : - First level decomposition: PSNR = 42.724 db with proposed scheme vs. PSNR = 42.400 db with Tao and Eskicioglu’s method. - Second level decomposition PSNR = 42.701 db with proposed scheme vs. PSNR = 42.230 db with Tao and Eskicioglu’s method. (a) Goldhill cover test image (b) Watermark used to test robustness against attacks. (c) Watermark used for rewatermarking attack. (d) Watermark used to compare the proposed scheme with Tao and Eskicioglu’s scheme. Fig. 2. Test platform Fig.3 . The PSNRs of twelve watermarked gray scale images each of size 512*512. (a) Watermarked image at first level decomposition, PSNR = 42.724 db. (b) Watermarked image at first level decomposition, PSNR = 42.723 db. Fig. 4. Watermarking results. First level decomposition Second level decomposition LL: SR = 1.000 HH: SR = 1.000 LL: SR = 1.000 HH: SR = 1.000 Fig. 5. Watermark extraction results. 0 10 20 30 40 50 PSNR (db) 3 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 60 61 Figure 5 provides the extraction results without applying any attack on the watermarked image. To evaluate the scheme robustness, we have applied different attacks to the watermarked Goldhill image. 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 suggest that: - The LL sub-bands are most resistant to lossy compression, filtering, geometrical deformations and noise addition. - The HH sub-bands are robust to nonlinear deformations of the gray scale. - Both sub-bands are resistant to rewatermarking. - Robustness is enhanced for second level decomposition. In particular, the visual quality of LL extracted watermarks and their SR values have been visibly increased in fig. 7 for the lossy compression, low-pass filtering, sharpening and noise addition. Comparison with previous methods In this part, we compare the experimental results of the proposed method with Tao and Eskicioglu’s method and Yuan’s method [13]. These two methods are based on the multiple watermarking approach in the DWT domain. Results are shown in figures 8, 9, 10, 11, 12 and 13, they are based on applying the same attacks on the Goldhill test image watermarked with the same binary logo for each comparison. 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 previous methods. In particular, the robustness of the LL sub-band has improved significantly for the gray scale deformation attacks such as histogram equalization. Figures 9, 11 and 13 provide the SR values 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 to the gray scale cover image and modifying the LL and HH sub-band coefficients in order to insert the binary watermark 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 embedding with second level decomposition results in better robustness. Objective evaluation shows that the proposed method outperforms Tao and Eskicioglu’s scheme in terms of fidelity and robustness. The proposed watermarking method can be further improved by automating the selection of the optimal thresholding parameter and appropriate scaling factor 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 = 0.920 LL: SR = 0.880 HH: SR = 0.880 Fig. 6. Extracting results for first decomposition level. 4 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 60 61 JPEG Compression Q=25 JPEG Compression Q=50 JPEG Compression Q=75 Gaussian filtring (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 = 0.926 LL: SR = 0.885 HH: SR = 0.885 Fig. 7. Extracting results for second decomposition level. Fig. 8. LL sub-band robustness comparison between 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) Histogram equalization, (i) Intensity adjustment ([0 0.8][0 1]), (j) Gamma correction (1.5), (k) Rewatermarking. Fig. 9. HH sub-band robustness comparison between Tao’s method and the proposed method for first level decomposition. (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. 0,1 0,3 0,5 0,7 0,9 a b c d e f g h i j k Tao and Eskicioglu's method Proposed method 0,1 0,3 0,5 0,7 0,9 a b c d e f Tao and Eskicioglu's method Proposed method 5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 60 61 Fig. 10. LL sub-band robustness comparison between Tao’s method and the proposed method for second 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) Histogram equalization, (i) Intensity adjustment ([0 0.8][0 1]), (j) Gamma correction (1.5), (k) Rewatermarking. Fig. 11. HH sub-band robustness comparison between Tao’s method and the proposed method for second level decomposition. (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. Fig. 12. LL sub-band robustness comparison between Yuan’s method and the proposed method for first level decomposition. (a) JPEG compression (Q=75), (b) Gaussian filtering (3*3), (c) rescaling (512->256->512), (d) Gaussian noise ([0 0.001]), (e) Gamma correction (1,5), (f) Cropping. Fig. 13. HH sub-band robustness comparison between Yuan’s method and the proposed method for first level decomposition. (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 survey of digital image watermarking techniques,” Industrial Informatics, 3rd IEEE International Conference on, pp. 709-716, 2005. [2] S. P. Mohanty, “Digital watermarking : a tutorial review”, unpublished. [3] E. Ganic and A. M. Eskicioglu, “Robust DWT-SVD domain image watermarking : embedding data in all frequencies,” Proceedings of the 2004 workshop on Multimedia and Security, pp. 166-174, 2004. [4] P-Y Chen and H-J Lin, “A DWT based approach for image steganography,” International Journal of Applied Science and 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 Computing, 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 synchronization using watermark key structure,” Security, Steganography and Watermarking of Multimedia Contents, pp. 536-547, 2004. [8] P. Bas, B. Roue and J-M Chassery, “Tatouage d’images couleur additif : vers la selection d’un espace d’insertion optimal,” Coresa03, 2003. [9] S. Rastegar, F. Namazi, K. Yaghmaie and A. Aliabadian, “Hybrid watermarking algorithm based on singular value decomposition and radon transform,” International Journal of Electronics and communication (AEÜ), pp. 658-663, 2011. [10] W. Bender, D. Gruhl, N. Morimoto and A. Lu, “Techniques for data hiding,” IBM Systems Journal , vol. 35, pp.313, 1996. [11] M-R Keyvanpour and F. Merrikh-Bayat, “Robust dynamic watermarking in DWT domain,” Procedia Computer Science vol. 3, 2010. [12] P. Tao and A. M. Eskicioglu, “A robust multiple watermarking scheme in the discrete wavelet transform domain,” Internet Multimedia Management Systems V, Proceedings of SPIE, pp. 133-144, October, 2004. [13] Y. Yuan, D. Huang and D. Liu, “An integer wavelet based multiple logo-watermarking scheme,” Proceedings of the First International Multi-Symposiums on Computer and Computational Sciences, vol. 2, pp.175-179, 2006. 0,1 0,3 0,5 0,7 0,9 a b c d e f g h i j k Tao and Eskiciolu's method Proposed method 0,1 0,3 0,5 0,7 0,9 a b c d e f Tao and Eskicioglu's method Proposed method 0,1 0,3 0,5 0,7 0,9 a b c d e f Yuan's method Proposed method 0,1 0,3 0,5 0,7 0,9 a b c d e f Yuan's method Proposed method 6