A Practical Print-and-Scan Resilient Watermarking for High Resolution Images

Transcription

1 A Practical Print-and-Scan Resilient Watermaring for High Resolution Images Yongping Zhang, Xiangui Kang 2, and Philipp Zhang 3 Research Department, Hisilicon Technologies Co., Ltd, Beijing, China. zhangyongping79@huaei.com. 2 School of Infor. Sci. and Tech., Sun Yat-Sen University, Guangzhou, China. issxg@mail.sysu.edu.cn 3 Research Department, Hisilicon Technologies Co., Ltd, Plano, Texas, USA. pzhang@huaei.com Abstract. Detecting the atermar from the rescanned image is still a challenging problem, especially for high resolution image. After printing and scanning, an image usually suffers global geometric distorts ( such as rotation, scaling, translation and cropping), local random nonlinear geometric distortions hich is simulated in Random Bending function of Stirmar, as ell as nonlinear pixel value distortions. The combination of both of the latter distortions is called nonlinear distortions. Many atermaring techniques, including the pilot-based atermaring techniques, are robust against the global geometric distortions but sensitive to the nonlinear distortions. Local random nonlinear geometric distortions remain a tough problem for image atermaring. In the setting of print-and-scan process, it becomes severer as combining ith nonlinear pixel value distortions. This may defeat a atermaring scheme, especially for the print-scanning of high resolution image. This paper proposes an effective pilot-based atermaring algorithm for the print-and-scan process. After careful analysis of the print-and-scan process, the pilot signal and the atermar are embedded and detected in the donsampled lo resolution image to deal ith the combination of local random nonlinear geometric distortions and nonlinear pixel value distortions. Theoretical analysis and experimental results demonstrate the proposed algorithm is robust to the print-and-scan process and at lo computational complexity. The major contribution of this paper is that e analyze the impact of nonlinear distortions in print-and-scan process and propose an effective atermaring scheme to conquer it. To our best noledge, our or also first addresses the issues of print-and-scan resilient atermaring for high resolution images. Keyords: Print-and-scan, atermar, non-linear distortions, robustness Introduction The print-and-scan process is commonly used for image reproduction and distribution. Print-and-scan resilient data hiding provides a viable authentication mechanism via

2 the multi-bit atermar hidden in a picture in the document. Document authentication of ID card, passport, driving license, image publication etc. is important today as the security concerns are higher than ever before. For example, the authentication of South China Tiger Image dre extensively concerns in China, even in the hole orld. But print-and-scan resilient data hiding has not been extensively researched. Many of them focus on detecting atermars []. The rescanned image is usually changed both in geometric and pixel value, and in both linear, non-linear ay. The atermar can be detected from the rescanned image only if the atermar can resist the combination of these distortions. Many atermaring techniques robust to geometric distortions are presented. One category is to embed the atermar in the global geometric transform invariant domain [][2]. The computational complexity of this technique is high, especially for high resolution images. Besides, the non-uniform log-polar mapping adopted in this technique might cause severe image fidelity problem. Another category [3][4][5] is to embed a pilot multiple times in the image at different spatial locations. The autocorrelation function of the atermared image ill then yield a pattern of peas corresponding to the embedded locations. Changes in this pattern of peas can be used to estimate the global geometric distortions that the atermared image has undergone. These methods have significant potential, but, similar to the above methods, do not consider the effect of the nonlinear distortions including nonlinear geometric distortions and nonlinear pixel value distortions. The atermar can be detected based on the pilot signals. But after printing-and-scanning, the nonlinear distortions destroy the periodicity of the pilot signal and the autocorrelation of the rescanned image ill not yield a pattern of peas. After careful analysis of the print-and-scan process, our proposal employs a donsampling operation in the atermar embedding and extraction process that simulate the atermaring embedding and extraction in the lo-resolution, thus reduce the effect of nonlinear distortions and mae sure that the pilot signal can be detected correctly. We can get the parameters of the global geometric distortions ith the help of pilot signal and invert the rescanned image before applying the atermar extractor. In this paper, e begin ith the analysis of the print-and-scan process in Section 2. Then e propose a novel atermaring algorithm in Section 3. Experimental results are presented in Section 4. In Section 5, e mae a summary. 2 Nonlinear Distortion Problems to Watermaring and Solutions In this section, e first mae a careful analysis on the print-and-scan process. Here e present ho the print-and-scan process affects the atermar. Then e propose the folloing solutions.

3 2. Analysis on the Print-and-Scan Process The process of print and scan is a complex composite of various attacs, hich cause various distortions including geometric distortion and pixel value distortion. The original image is transmitted into control system of the printer. Before Raster Image Processor (RIP) in the control system converts the original image into halftone image and corresponding electronic pulse level, the printer ill mae the Dots Per Inch (DPI) value of image equal to the physical DPI value of the printer. The electronic pulse level is converted into optical level and forms latent image on Organic Photo Conductor (OPC) drum of electrophotographic system. After development process, transferring process and fusing process, latent image is transformed into hardcopy output. In the scanning process, a pure light emitted to the hardcopy global. Charge Coupled Device (CCD) sensors receive the luminance of the light and convert it into the electronic signal. At last, the generated digital image generated is transmitted to PC via Graphic User Interface (GUI). [7] According to the process of the print-and-scan process, e can find that both the printing process and scanning process introduce geometric distortions. The distortions during these to processes are different. Converting the electronic pulse into optical level and hardcopy process introduces the nonlinear geometric distortions. Consider an image I, and a nonlinear distorted version of this image, I. Then e can rite that D( = I '( I( () here D( represents the difference of pixel value beteen I and I. Let x and y represent the magnitude of nonlinear geometric distortion in the horizontal direction and in the vertical direction respectively. The position of pixel I( is translated to the position ( x + y + in the rescanned image. The nonlinear adjustment in the printer and scanner also introduce the distortion on the value of pixel. The nonlinear adjustment, hich is called gamma teaing, is performed in the printer to mae sure the printed images appear the same as on a monitor: γ p B = I( (2) here B represents the actual light brightness. When the image is scanned, it must be compensated to mae sure it loos fine to us hen vieed on a monitor: γ s I' ( = B (3) So, I γ p γ s ' ( = I ( (4) Besides these distortions, there is some random noise in the print-and-scan process. And let N represent the random noise. Then e can get γ s D y = I'( I( = I( x y + N I( γ p (5)

4 When γ > γ, this nonlinear adjustment amplify the effect of nonlinear p s geometric distortion. And D is distributed in the hole image. It may become large enough to affect the atermar extracting. There is one measure to deal ith this problem if the resolution of image is lo. We can scale the atermared lo resolution image up to a higher resolution before printing it out. And in the scanning process, the printed atermared image is donsampled to a lo resolution as the original resolution. This operation can improve the robustness against the nonlinear distortions. Some experiments on Digimarc atermaring tool in Photoshop are performed. In our experiments, 00 images ith 800-by-600(4:3) are utilized. The embedding strength is as large as possible under the limitation of invisibility and the PSNR value is 36.3dB averagely. It shos good performance hen e up-scale the atermared image to 4096-by-3076(4:3, hich is the size of A4) in the printing process. The Bit Error Ratio (BER) of atermar extraction from the rescanned atermared images is about 3.5%. But unfortunately, ith the development of the multimedia technologies, the resolutions of images become higher and higher in the application. Especially in the document printing, the driver of printer ill convert a text to an image. The resolution of this image is very high and based on the size of paper printed on. No the above measure does not or. For example, hen the atermared image hose resolution is 4096-by-3076 is printed on A4, e cannot up-scale the atermared image because of the paper scope. No the nonlinear distortions can destroy the atermar easily. Experiments on the Digimarc atermaring tool are performed on four images ith 4096-by-3076 or 3076-by The atermared images are printed ith the original resolution on A4 paper. No atermar can be extracted from the rescanned atermared images. Another problem is that many existing print-and-scan resilient atermaring algorithms are complex and time consuming. Many complex methods are included to mae the atermaring algorithm to be robust against the print-and-scan process. Computation efficiency is an important consideration in application. There is still another measure proposed in this paper to deal ith the above problems. We can adopt the method of embedding and extracting atermar at lo resolution shon as follos to get better robustness against nonlinear distortions and higher computation efficiency. 2.2 Proposed Solution Based on the above analysis and the experiments ith Digimarc atermaring tool on the print-and-scan process, our solution is to add a don-sampling operation to achieve the lo-resolution hich is necessary in the atermar embedding and extraction process. In atermar embedding, e first don-sample the original image ith and embed the atermar in the don-sampled image. Let I ( y ) represent one pixel in the don-sampled atermared image. We can obtain,

5 I ( I ( x ) + ( y ) + j) j= = 2 (6) ' Assume I ( to be the pixel of the atermared image and ( to be the pixel of the rescanned atermared image, here x = ( x ) + i j [, ] (7) y = ( y ) + j In atermar extraction, the rescanned image is also don-sampled to / of the original size. We can obtain, ' ( ( x ) + ( y ) + j) ' I j= = (8) 2 I The difference beteen Eq.(6) and Eq.(8), hich is the effect on the image at lo resolution from the nonlinear distortion, can be described as, here D = D i, j j= I j ' j ( I ( = = 2 D and ' ' D I ( I ( = I (( x ) + ( y ) + j) I (( x ) + ( y ) ) D 2 i, j + j I (9) =, hich is shon in Eq.(5). The impact of random noise N in Eq.(5) can be removed or lessened ith the summation operation in Eq.(9). Most importantly, the nonlinear 2 geometric distortion in a don-sampled rescanned image ill be reduced to / of the distortion in a rescanned image ithout don-sampling. If is great enough, the effect of nonlinear geometric distortion almost may be ignored. 3 Proposed Scheme The ey idea of our proposed scheme is to embed and detect the atermar in lo resolution. This ill reduce the effect of nonlinear geometric distortion and ensure the survival of atermar from the print-and-scan process. The diagram of our atermaring scheme is illustrated in Fig.. 3. Watermar Generation In our atermaring algorithm, there are to types of atermaring signal. One is pilot pattern p, hich is a hite noise pattern and embedded in certain blocs. Through the extraction of pilot patterns, the rescanned image can be resynchronized. The other is atermar payload { b i }. The atermar payload is often chosen to be orthogonal noise-lie patterns or spread spectrum modulated by a basis of n orthogonal noise-lie patterns {u i } via,

6 Fig.. Bloc diagram of the proposal scheme. (a) is atermar embedding process. (b) is atermar extraction process. here {0, } b or {, } i n = i= b [8]. i b i u i (0) 3.2 Watermar Embedding in Lo Resolution In order to resist the nonlinear geometric distortions introduced in the print-and-scan process, the lo-resolution image based atermar embedding method is proposed. The main steps are as follos. Step Don-sample the original image I to / of the original size to get I. In our experiments, e use =0. Step 2 Divide I into the blocs and perform Discrete Cosine Transform (DCT). Step 3 Embed the pilot pattern p and atermar pattern in DCT coefficients f of different blocs (see Fig. 2).

7 Fig. 2. The partition model. The pilot patterns are embedded in the blac blocs. And the atermar patterns are embedded in the other blocs. The gray regions are exhaustive searching indos. f ( y ) = f ( y ) + α () here α is the atermaring strength. We use α = 0. 6 in our experiments. And here p is a hite noise pattern and is a set of hite noise patterns. Here e use the popular linear, additive atermaring scheme [8] in the atermar embedding. Other embedding schemes, such as linear [0][] and nonlinear multiplication atermar scheme [2], may be used to achieve better performance in terms of robustness or invisibility of atermar. Step 4 Perform inverse DCT transform on the atermared DCT coefficients f of every bloc and get I. Step 5 Get the atermar mas as, Mas = I I (2) Step 6 Up-scale Mas ith to the original size and add it to the original image to get the atermared image I. 3.3 Watermar Extraction The atermar extraction procedure includes three steps. Step Don-sample the rescanned image at different ratio and detect the positions of four pilot patterns by exhaustive searching in the certain indos (see Fig. 2).

8 Assume R img to be the resolution of the original image and R scan to be resolution of the scanner. It is ell non that the difference beteen R img and R scan ill introduce scaling beteen the original image and the rescanned one. And the scaling parameter is R scan P scale = (2) Rimg In order to extract the pilot signal in the lo resolution, e also don-sample the atermared image ith /. According to Section 3.2, the effect of nonlinear geometric distortions can be reduced. So the possible ratio of don-sample is Rate R img = (3) R R scan is certain in the extraction process. And although R img is unnon to the detector, the possible value is limited in application. We ill exhaustive searching in all possible value of Eq.3. We can get the searching result at hich the correlation beteen the original pilot signal and the extracted one is highest. Don-sample the rescanned image ith that result ratio and find the locations of four pilot signals in the don-sampled rescanned image. Step 2 Get the parameters of global geometric distortions. With the help of four pilot signal, e can get rotation angles and four sets of parameters of translation [7]. We use the average values of them to resynchronize the don-sampled rescanned image. Because the positions of four pilot signals can be gotten correctly, e can get the exact orientation ith their help. Step 3 Decode the atermar payload bit by bit from the recovered image. The don-sampled image is divided into blocs. Select the DCT coefficients of blocs at the same positions as in the embedding procedure. Compute the correlation beteen u i and the selected coefficients. If the correlation is greater than a threshold T, then indicate that the bit of atermar b i is. scan 4 Experimental Results We evaluate the proposed algorithm using 4 high resolution images from Standard High Precision Pictures (SHIPP). The parameter of don-sample is set to 0. The don-sampled image is divided into blocs ith size 64-by-64 and there are 24 blocs in one test image. Except four blocs embedded the pilot signals in, e embed 64bits in each of the other tenty blocs. So the largest atermaring capacity is up to 280 bits. In our application, e embed the same atermar in these tenty blocs and the atermaring capacity is 64 bits. The PSNRs of the atermared images are presented in Table. From this table, e can find that visual quality of the atermared images is acceptable. And the atermared images are shon in Fig. 3.

9 Table. PSNRs of the atermared images. Image Bride Harbor Wool Bottle Resolution 3072-by by by by-4096 PSNR(dB) Fig. 3. Watermared images We printed the atermared image out by Ricoh Aficio 2032 and scanned the printed atermared images by HP scanjet 8250, both common commercial products. We employ different models of Ricoh laser printer and HP scanner to evaluate the robustness of our proposed atermaring algorithm. The resolution of Aficio 2032 is set to 600 dpi. And the scanning resolution is set to 300 dpi and 600 dpi. The detection results are shon in Table 2. The BER of extrcted 64 bits is less than.6%. So e can conclude that the proposed algorithm is robust to print-and-scan process. Table 2. The detection results BER(%) Image Scanning Resolution = 300 dpi Scanning Resolution = 600 dpi Bride 0 0 Harbor 0 0 Wool 0 0 Bottle.6 0 We also evaluate the proposed algorithm s robustness against JPEG compression. The atermared images are compressed under different JPEG quality factors. The quality factors are set to be from 50 to 90 and at intervals of 0. All atermars can be extracted correctly ithout any error bit from compressed images. In the experiments, the atermaring algorithm is implemented ith CPU PIV 2.4G, RAM 2G and Matlab 6.5. It taes about 7.89s for the embedding, and taes 26.08s for the extraction despite the very high resolution of images, so it is efficient in computation.

10 5 Conclusions In this paper, e analyze the impact of the combination of local random nonlinear geometric distortions and nonlinear pixel value distortions in print-and-scanning process, and a novel print-and-scan resilient image atermaring algorithm is proposed. The atermars are embedded in the don-sampled lo resolution image to reduce the effect of nonlinear distortions in printing process. And the pilot patterns are embedded to be resilient global geometric distortions in print-and-scan process. Theoretical analysis and the experimental results demonstrate that the proposed algorithm is robust to the print-and-scan process and its computational efficiency is high. Future or is to perform more testing on this algorithm by more test images. Acnoledgement This or as supported by NSFC ( ), NSF of Guangdong ( ), and NSF of Guangzhou (2006Z3-D304). References. C.Y. Lin, M. Wu, J. A. Bloom, I.J. Co M.L. Miller, and Y.M. Lui: Rotation, Scale and Translation Resilient Watermaring for Images, IEEE Transactions on Image Processing, vol. 0, no. 5, pp (200). 2. J. O Ruanaidh, and T. Pun: Rotation, Scale and Translation Invariant Spread Spectrum Digital Image Watermaring, Signal Processing, vol. 66, no. 3, pp (998). 3. P. C. Su, and C. C. Jay Kuo: Synchronized Detection of The Bloc-Based Watermar ith Invisible Grid Embedding, In: SPIE Photonics West, Security and Watermaring of Multimedia Contents III, San Jose, California (200). 4. M. Kutter: Watermaring Resisting to Translation, Rotation and Scaling, In: Proceedings of the SPIE: Multimedia systems and applications, Boston, USA (3528), pp: (998). 5. S. Voloshynovsiy, F. Deguillaume, T. Pun: Multibit atermaring robust against local nonlinear geometrical distortions, In: IEEE International Conference on Image Processing 200, Thessaloni Greece, pp: (200). 6. C. Y. Lin, S. F. Chang: Distortion Modeling and Invariant Extraction for Digital Image Printand-Scan Process, In: Int. Symposium on Multimedia Information Processing (999). 7. L. Yu, X. Niu, and S. Sun: Print-and-Scan Model and the Watermaring Countermeasure, In: Image and Vision Computing, 23, pp: (2005). 8. I.J. Co J. Kilian, T. Leighton, and T. Shamoon: Secure Spread Spectrum Watermaring for Multimedia, IEEE Transactions on Image Processing, vol. 6, no. 2, pp: (997). 9. K. Solan U.Madho, B.S.Manjunath, S. Chandrasearan, and I. El-Khalil: Print and Scan Resilient Data Hiding in Images, IEEE Transactions on Information Forensics and Security, vol., no. 4, pp: (2006).

11 0. Qiang Cheng and Thomas S. Huang: Robust Optimum Detection of Transform Domain Multiplicative Watermars, IEEE Transactions on Signal Processing, vol.5, no. 4, pp: (2003).. Xiangui Kang, Xiong Zhong, Jiu Huang, and Wenjun Zeng: An Efficient Print-Scanning Resilient Data Hiding Based on a Novel LPM, In: IEEE International Conference on Image Processing 2008, San Digeo, CA, USA (2008). 2. Wei Liu, Lina Dong, and Wenjun Zeng: Optimum Detection for Spread-Spectrum Watermaring That Employs Self-Masing, In: IEEE Transactions on Information Forensics and Security, vol. 2, No. 4, pp: (2007). 3. D. He, and Q. Sun: A Practical Print-Scan Resilient Watermaring Scheme, In: IEEE International Conference on Image Processing 2005, vol., pp: (2005).