SIGNIFICANT PIXEL WATERMARKING IN CONTOURLET OMAIN Jayalakshmi M., S. N. Merchant, Uday B. esai SPANN Lab, Indian Institute of Technology, Bombay email: jlakshmi, merchant, ubdesai @ee.iitb.ac.in Keywords: Contourlets, Pyramidal irectional Filter Banks, Laplacian Pyramid Abstract In this paper, we propose a robust digital watermarking scheme for images in contourlet domain. Conventional methods embed multiple copies of the watermark, whereas, we try to locate the best coefficients to hide a single copy of the watermark and achieve robustness. We propose the concept of significant pixels in contourlet domain and explore the same for hiding the watermark. We have carried out experimentation with both gray and color images containing lot of curves and texts like maps and those with textures. To test the robustness, the watermarked images were subjected to different attacks which distort the watermark while leaving the image with good perceptual quality. The proposed method outperformed conventional techniques using WT, CT and Hadamard transform. A comparison of the proposed method with two other watermarking approaches in contourlet domain is also performed to show that significant pixel watermarking is indeed a suitable choice. A distinct and remarkable performance was observed in images like maps which contain lot of texts and curves. 1 Introduction igital watermarking primarily means inserting a copyright information into a cover work and is proposed as a solution to illegal copying and tampering of the original data. The existing transform domain techniques using CT, FT and WT locate regions of high frequency or middle frequency to embed information imperceptibly [8, 9]. Contourlet transform gives a multiresolution, local and directional expansion of images using Pyramidal irectional Filter Bank(PFB) [7]. Eventhough wavelets are universal in representing point singularities, they are inefficient to represent discontinuities which frequently appear in two and higher dimensions. Curvelets were introduced to overcome this drawback of wavelets [4]. Contourlet transform was proposed as an improvement on curvelet transform using a double filter bank structure. The PFB combines Laplacian Pyramid(LP) with a directional filter bank [1, 3]. The former captures point discontinuities and the latter links these into linear structures. In this paper, firstly we propose the concept of significance factor in contourlet domain. This measure is defined depending on the multiresolution dependency of the contourlet transformed coefficients. Secondly, we recommend to use significant pixels for hiding the watermark. The proposed method embeds a single copy of the watermark and still achieves good robustness against conventional methods. Hence we prove that the choice of pixels depending on significance factor is indeed better than their choice without considering the dependencies across scales. The watermarked images by the proposed method are of very good visual quality irrespective of the nature of the images selected. The watermarked gray images are subjected to three different types of attacks, namely, mean filtering, quantization and JPEG compression. Color images are subjected to Gaussian noise, salt-pepper noise with median filter and other intentional attacks using Adobe Photoshop software also. For the purpose of comparison we have simulated five other methods: two of them in contourlet domain itself and one each in WT, CT and Hadamard transform domains. The normalized correlation coefficients after different attacks were calculated for all the methods and significant pixels in contourlet domain performed exceptionally well in images like maps which contain lot of texts and curves. The main contribution of the paper is in locating the best coefficients to hide a single copy of the watermark with robustness. The paper is organized as follows. In Section 2, we propose the concept of significance factor in contourlet domain. Section 3 describes the algorithm for watermark embedding using the concept from Section 2. Experimental results and observations are included in Section 4. Conclusions are drawn in Section 5. 2 Concept of significance factor in contourlet domain One of the unique properties of contourlet transform is that we can specify the number of directions in the bandpass images at every scale. The directional decomposition used in our method is shown in Figure 1. Here the number of directional bands doubles at every scale of multiresolution pyramid. From Figure 1 it is clear that L is the low pass band and W, X, Y and Z are the directional bands. Each directional subband is defined with subscript, e.g., the band Y is divided into eight subbands named as. The binary watermark selected is of size and hence we locate 256 pixels of visual importance from the coarsest bandpass image W, chosen for embedding.
X = = > > Y ; ;??? 3 Watermark embedding Significance factor of each of the pixels in band W is calculated as described in the previous section and these values are then sorted. The pixels with highest significance factor will be chosen for watermark insertion. Let W(x,y) be a significant pixel selected for watermarking from band W and dfe_g be the multiplication factor used for embedding. We perform additive watermarking to get the watermarked pixel W (x,y) [6]. $ g h [ik! '$ h jik &]e k A (5) where k A is the watermark bit obtained after randomizing the binary watermark selected. Additive watermarking is performed on all the images and the multiplication factors chosen are 35 and 50 for gray level and color images respectively. Nevertheless, it is possible to quantize pixels differently, depending on human visual system models [2]. Figure 1: Contourlet decomposition To select a pixel with high visual importance, we calculate the significance factor S(i,j) for each pixel W(i,j), depending on the multiresolution dependency, as follows.! #" $ "&'" (*) +"&," -.) +"&," /!) " (1) Here, " ( ) +", " - ) +" and " / ) 0+" are the sum of absolute values of all children coefficients of $ 01 in band X, Y and Z respectively. To see how the children coefficients are located let us a consider a pixel 2 from band 2. Let the sum of absolute values of children of this pixel in band X be denoted by 354067 0. They are selected from bands 38 and 39 as follows. >? > E FHG ( )0: ; 0< " ( 0+" (2) @ ; A @CB Children of each pixel 3I 0 of subband 3J, are selected from bands! K L +M NM L NM O and MQP. Let the sum of absolute values of these children coefficients from Y be 4SR+ and Z be MT4SRN. These are calculated as follows. -7)0UE; /!)0UE; 0< 0< >? @ ; A @CB >? @ ; A @CB " - = " / = V ; V ; [Z\ FWG FWG +" (3) +" (4) The same method is followed to find the parent child dependencies and significance factor in the band 2] also. But here children coefficients will be selected vertically from bands 35L 3_^. O to and Ma` to Ma cb. Pixels with high significance factor will be selected for watermark embedding. For retrieving the watermark we need a copy of the original image and hence the proposed algorithm is non-blind. Subtracting the original image pixels from the watermarked pixel in the transform domain, we get the scaled watermark. 4 Experimental results Watermarking was performed on both color and gray images. Here we discuss the results with gray images. The results with color images are discussed separately in this section. We have experimented with seven gray images: four of them are maps as depicted in Figure 2. The other three are Lena, Barbara and Baboon. Since these are well known images, due to space limitations, we have not included them in this paper. However, all the simulations were performed on these test images also and the results were observed. For the purpose of comparison, we have included the results of five other methods, each selecting the 256 highest absolute valued coefficients for embedding. The first and second methods, referred to as Method 1 and Method 2 respectively, embed watermark in contourlet domain. Method 1 follows same decomposition as in Figure 1 and selects coefficients from band W. Method 2 follows the curve scaling relation for curvelets for decomposition and chooses the coarsest bandpass image for embedding [4]. Here the number of directional subbands doubles at every other scale of decomposition. The third method is based on discrete wavelet transform and selects coefficients from the detail bands of four level decomposed image for watermark embedding. The fourth and fifth methods also select the highest absolute coefficients from CT and Hadamard transformed images respectively [6]. Both these methods exclude the C value while embedding. The watermarked images corresponding to Image 3 by the proposed method and the other five methods are included in
(a) Image 1 (b) Image 2 (a) Proposed (b) Method 1 (c) Image 3 (d) Image 4 (c) Method 2 (d) Wavelet Figure 2: Original test images Figure 3. The watermarked images are perceptually very similar to the original. To keep the visual distortions to minimum we calculate the Peak Signal to Noise Ratio (PSNR) of the watermarked image. With the chosen e factor the PSNR is found to be approximately 47dB and above for all the test images. (e) CT (f) Hadamard 4.1 Resilience to attacks Figure 3: Watermarked images of Image 3 Watermarked images should be resistant to those attacks which would retain the visual quality of the watermarked image and would try to distort the embedded message. Since maps include text and curves, conventional signal processing operations, like Gaussian noise addition and median filtering, distort the image considerably. In this paper we have considered only those attacks which do not distort the watermarked image. In particular, we consider attacks like mean filtering, quantization and JPEG compression. The correlation coefficient (l ) between the original and retrieved watermark has been chosen as the measure of similarity. The correlation coefficients after mean filtering the watermarked images are tabulated in Table 1. The original watermark and retrieved watermarks after mean filtering from Image 3 by different embedding methods are shown in Figure 4. From the percertual quality of the retrieved watermarks it is obvious that the proposed method outperforms other methods. Image Proposed Method1 Method2 Wavelet CT Hadamard Image1 0.88 0.85 0.88 0.52 0.68 0.27 Image2 0.96 0.93 0.88 0.78 0.95 0.64 Image3 0.91 0.85 0.80 0.48 0.78 0.34 Image4 0.81 0.59 0.59 0.63 0.52 0.24 Lena 0.95 0.95 0.85 0.95 0.91 0.65 Barbara 0.91 0.85 0.91 0.92 0.91 0.61 Baboon 0.95 0.99 0.91 0.97 0.81 0.67 Table 1: Correlation coefficient after mean filtering Another signal processing operation considered as an attack is the quantization of the watermarked images. Two different cases of quantization of the watermarked images were experimented. The watermarked pixels were quantized to multiples of 10 and multiples of 20. Figure 5 shows the retrieved watermarks after quantization of the watermarked pixels to multiples of 20 by all the methods considered. From
Z n p u (a) Original watermark(spann) (a) Proposed y (b) Method 1 lmtn fuez lmonkzuexwzwz (c) Method 2 l{on uep1u (b) Proposed (c) Method 1 (d) Method 2 lmonk p lqon r1s lmtn fuvp (d) Wavelet lqon fuvxvzyz (e) CT lq,nkfswsvzu (f) Hadamard l{on y Figure 6: Retrieved watermarks after JPEG compression with Q =10 from Image 3 (e) Wavelet lmonk Zu w (f) CT lqon fuvr (g) Hadamard lmtn xyxyswp 4.2 Color Images Figure 4: Original watermark and retrieved watermarks after mean filtering Image3 the visual similarity of the retrieved watermarks with the original, the robustness of the proposed method is obvious. Watermarking of color images is usually performed in the luminance components of any selected color space representation and we have used the YCbCr representation for watermark insertion [5]. Here the luminance component alone is considered while locating the significant pixels of the images and the selected significant pixels are watermarked by additive embedding. But we have used a multiplication factor of 50 and the PSNR was found to be 44dB and above for all the images. Figure 7 shows three selected color images and their respective watermarked images. It is obvious that the watermarked images are not perceptually distorted with respect to the original image. (a) Proposed lmonk rvz1xwr (b) Method 1 lqon r nwx (c) Method 2 lqon fuwsvnyn 4.3 Resilience to attacks in color images (d) Wavelet lmonk r1uwsvn (e) CT lqon fu nwp (f) Hadamard lmtn In the case of color images, we performed various attacks like mean filtering, compression, Gaussian noise addition, salt and pepper noise with median filter, sharpening the image edges and increasing the graininess of the image. These attacks were tried since they did not distort even the maps too much as in the case of gray images and a few relevant results are shown. Figure 5: Retrieved watermarks after quantization of Image 1 to multiples of 20 JPEG compression is one of the important attacks which any image watermarking algorithm should be resistant to. For the purpose of comparison, we have included the retrieved watermarks from Image 3 after JPEG compression with Quality factor(q) 10, in Figure 6. Figure 8 shows the retrieved watermarks from the watermarked images after mean filtering attack. Retrieved watermarks in Figure 8a, 8b, and 8c are by the proposed method and 8d, 8e, and 8f are by wavelet based method from the three images considered. From the correlation of the retrieved watermarks after mean filtering with the original watermark, we can assure that the proposed method works better than the wavelet based method in case of images like maps. The effect of salt-pepper noise with median filter and Gaussian noise addition were also considered in the case of color images.
(a) From peppers (b) From Map w 1 (c) From Map 2 (d) From peppers (e) From Map 1 (f) From Map 2 lmtn r r (a) Peppers lq,nk pwp l{ (b) Watermarked Image lmtn pyrvzyz lq,nk p wr r l{tnk r r Figure 8: Retrieved watermarks after mean filtering: a, b, and c - by proposed method ; d, e and f - by wavelet based method (c) Map 1 (d) Watermarked Image measure, was calculated after different types of attacks with varying intensities. The results show the distinct advantages of selecting the contourlet domain significant pixels over other methods. 5 CONCLUSION (e) Map 2 (d) Watermarked Image Figure 7: Original and watermarked images Both salt-pepper noise and Gaussian noise added have zero mean and variances 0.01 and 0.001 respectively. Figure 9 shows the images of the two colored maps after salt-pepper noise addition and median filtering. The retrieved watermarks from these images are also shown in Figure 9c and 9d. Figure 10 shows the Gaussian noise added images and the retrieved watermarks from them. The attacks, namely, sharpening and increasing the grainiess were performed using Adobe Photoshop software. Graininess added has intensity 40 and contrast 50. The images in Figure 11 and Figure 12 show the attcked images and retrieved watermarks after graininess addition and sharpening respectively. In short, significant pixels defined in contourlet domain proved extremely suitable for robust invisible watermarking. The normalized correlation coefficient, chosen as the similarity The significance factor defined in contourlet domain, based on the multiresolution dependency of the contourlet coefficients, was proposed for image watermarking. The significant pixels were found better than the highest absolute coefficients in different transform domains while a single copy of the watermark is embedded. The proposed algorithm shows significant improvement over conventional methods, especially for images like maps which contain lot of texts and curves. This technique is currently being improved as a blind technique by the authors. References [1] R. H. Bamberger and M. J. T. Smith. A filter bank for the directional decomposition of images: theory and design. IEEE Trans. on Signal Processing, 40:882 893, April 1992. [2] Mauro Barni, Franco Bartolini, and Alessandro Piva. Improved wavelet based watermarking through pixel-wise masking. IEEE Trans. on Image Processing, 10(5):470 477, May 2001. [3] P. J. Burt and E. H. Adelson. The laplacian pyramid as a compact image codes. IEEE Trans. on Communications, 31:532 540, April 1983. [4] Emmanuel J. Candes and avid L. onoho. Curveletsa surprisingly effective nonadaptive representation for objects with edges. Saint-Malo Proceedings, 1999.
l{tnk p wn p lq,nk rwpyn Figure 9: Salt-pepper noise and median filtered image and retrieved watermarks lm,nk pyswx l{tnk r1uwsvn lqon rypwrwz Figure 11: Grain images and retrieved watermarks lq,nk rwx1svp lm,nk pwrwzwz lqon p yn p Figure 12: Sharpened images and retrieved watermarks Figure 10: watermarks Gaussian noise added images and retrieved representation. IEEE Trans. on Image Processing, 14:2091 2106, ecember 2005. [5] Chun-Hsien Chou and Kuo-Cheng Liu. An oblivious and robust watermarking scheme using perceptual model. 4th EURASIP Conference Video/Image Processing and Multimedia Communications, pages 713 719, July 2003. [8] C. T. Hsu and J. L. Wu. Multiresolution watermarking for digital images. IEEE Trans. on Circuit and Systems, 45:1097 1101, August 1998. [6] I. Cox, J. Kilian, F. T. Leigton, and T. Shamoon. Secure spread spectrum watermarking for multimedia. IEEE Trans. on Image Processing, 6:1673 1687, ecember 1997. [9] eepa Kundur and imitrios Hatzinakos. Towards robust logo watermarking using multiresolution image fusion principles. IEEE Trans. on Image Processing, 6(1):185 198, February 2004. [7] Minh N. o and Martin Vetterli. The contourlet transform: An efficient directional multiresolution image