Image Splicing Detection Based on Texture Consistency of Shadow

Transcription

1 Image Splicing Detection Based on Texture Consistency of Shadow 1 Yongzhen Ke, Weidong Min, 3 Xiuping Du, 4 Dandan Li 1,,4 School of computer science and software engineering, Tianjin Polytechnic University, China, keyongzhen@tjpu.edu.cn *3 School of internet education, Tianjin University, China, duxiuping@tju.edu.cn Abstract This paper takes the advantage of the property that the shadow will not obviously change the surface texture of object and presents a method based on texture feature consistency of shadow for image splicing detection. First, texture features of shadows areas and its adjacent lit areas are extracted. Then, the Euclidean distance is used to measure the similarity of texture features. Last, a texture features fusion framework integrating multiple texture cues for forgery detection is proposed to improve algorithm performance, robustness, and adaptability. Experimental results show that the algorithm is simple and effective. 1. Introduction Keywords: Image Forensic, Texture, Shadow, Feature Fusion Nowadays, digital cameras are playing an important role in our daily life as well as in forensic affairs. However, the sophisticated digital image processing software, such as Photoshop and ACDSee, have made image tampering much easier to operate but harder to detect. Today, splicing operation is widely used under various conditions. Especially in recent years, numerous photos in photography competition are exposed to be synthesized. This urges us to find a ways to distinguish the authentic and tampered photos. In order to solve this problem, digital image forensic technology emerges at a historic moment. As a digital image forensic technology, blind forensic technology is becoming a new hotspot in the field of multimedia security with a wide application prospect because of its advantages of identifying image authenticity and source without relying on any signature extraction or pre-embedded information. At present, the detection algorithm for image splicing can be roughly divided into two categories: detection algorithm based on authenticity of the local area and detection algorithm based on source inconsistency. The former contains many detection algorithms. Farid [1] used higher-order spectral analysis to analyze the high order coefficient generated by splicing operation, after that Ng et al. [] used bicoherence along with other features to detect the high order correlation introduced by splicing. Hilbert-Huang transformation, Discrete Cosine Transform and wavelet decomposition were also used in literature [3-5]. Lipsiz law [6] and phase consistency method [7] can be used to measure the signal smoothness or sharpness according to wavelet coefficients. Hui Fan et al.[8] proposed a blurred image tampering detection method using hommonorphic filtering in dyadic contourlet domain to enhance the blurred edges. Some algorithms also used the illumination direction [9, 10] and image chroma [11] to detect the image. Approaches based on the consistency of imaging source are based on the fact that natural images are usually obtained through data acquisition devices, which introduce uniform characteristics to the entire image, and henceforth the variation in the local characteristics across the image can be used to detect tampering. The characteristics of the lens [1], sensor pattern noise [13], color filter array interpolation [14-16], and consistency of camera response function [17, 18] have been used to detect image forgeries. Shadows provide important visual cues for depth, shape, contact, and lighting in our perception of the world. Shadow as an important feature of the digital image, has already been used in image splicing detection [19, 0]. Methods detecting photographic composites based on shadow geometry and shadow photometry are introduced in literature [19]. Qiguang Liu [0] proposed a framework for detecting tampered digital images based on photometric consistency of illumination in shadows. The approach proposed in this paper is based on both the assumption that the shadow as well as main body is copied and pasted from another image during forgery and the property that the shadow will not obviously change the surface texture of object. In other words, the shadow areas with their adjacent non-shadow area should have the same or similar texture. So, the texture in the shadow area is Journal of Convergence Information Technology(JCIT) Volume8, Number4,Feb 013 doi: /jcit.vol8.issue4.101

2 not consistent with that in the original lit area in tamper image. Based on the above mentions, this paper presents a method based on texture feature consistency of shadow for image splicing detection. First, we extracted texture features of shadows areas and its adjacent lit areas. Then, the Euclidean distance is used to measure the similarity of texture features. Last, texture features fusion is used to improve performance, robustness, and adaptability. This paper is organized as follows: Section describes the proposed algorithm in details. The valuable experiment results are shown in Section 3. Section 4 concludes this paper.. The proposed image splicing detection algorithm based on texture consistency of shadow.1. The principle of texture consistency of shadow Shadows contain a wealth of information in digital images. There are many important visual cues of the shadow [1, ] : a) Shadow is dark but does not significantly change either the color or the texture of the background covered. b) Shadow is always associated with the object that casts and the behaviour of that object (e.g. if a person opens his arms, the shadow will reflect the motion and the shape of the person). c) Shadow shape is the projection of the object shape on the background. For an extended light source (not a point light source), the projection is unlikely to be perspective. d) Both the position and strength of the light source are known. e) Shadow boundaries tend to change direction according to the geometry of the surfaces on which they are cast. Our approach takes the advantage of the property that the shadow will not obviously change the surface texture of object. As shown in Figure 1, the shadow area (S) of the real scene should have the same or similar texture features with adjacent non-shadow area (N). We extract LBP texture feature (see figure ) and difference matrix texture feature (see figure3) of digital images for non-shadow area and shadow area. Figure 1. Images with shadow where N represents the non-shadow area and S represents the shadow area (a) (b) (c) (d) (a) LBP of non-shadow area of left image in Figure 1.(b)LBP of shadow area of left image in Figure 1. (c) LBP of non-shadow area of right image in Figure 1.(d)LBP of shadow area of right image in Figure 1. Figure. Local Binary Pattern (LBP) of different image for non-shadow area and shadow area

3 (a) (b) (c) (d) (a) Difference matrix of non-shadow area of left image in Figure 1. (b)difference matrix of shadow area of left image in Figure 1. (c) Difference matrix of non-shadow area of right image in Figure 1.(d)Difference matrix of shadow area of right image in Figure 1. Figure 3. Difference matrix of different image for non-shadow area and shadow area Figure and Figure 3 show that shadow area have the same or similar texture to its adjacent nonshadow area on the same image, called texture consistency of shadow, but textures of shadow area and non-shadow area between the different images are inconsistent.. Image splicing detection algorithm During image forgery, the shadow as well as main body is copied and pasted from another image. Based on the above characteristics, the texture of shadow area is inconsistent with that of non-shadow area for Mosaic image. Image splicing detection algorithm based on texture consistency of shadow is shown in figure 4. First, the shadow area and adjacent non-shadow area are extracted, and then texture features are obtained. Last, Calculation of similarity between the textures is applied to decide whether the input image is an original image or a forgery image..3 Texture features Figure 4. The image splicing detection algorithm A number of texture features such as Gray-level Co-occurrence Matrix, Local Binary Pattern, Gabor Filtering, and Difference Matrix are used in the proposed algorithm Gray-level Co-occurrence Matrix One of the most popular and powerful ways to describe texture is using of GLCM [3]. It represents an estimate of the probability that a pixel has a gray level intensity g i and a neighboring pixel has an intensity g j, where g j [0, Ng 1] and Ng is the number of available gray levels in the image. For the brightness level b 1 and b, co-occurrence matrix C is as follows: N N P b P ' ' b C 1 (1) b1, b xy xy x 1 y 1 where x ' is given by the distance d and offset angle as follows :

4 and the same as ' y ' x x d d d cos ( 1, max ) 0, ' y y d d d sin ( 1,max ) 0, (3) The formula (1) can be applied to the image to get a symmetric square matrix. In the gray-level cooccurrence matrix generation process, the maximum distance is set to a pixel, and the direction is set to select each point of the four nearest neighbor. GLCM is based on a point and its four nearest points obtained symbiotic relationship. GLCM reflects the direction, adjacent intervals and change of information of gray image..3.. Local Binary Pattern Local Binary Pattern was introduced by Ojala in 1996 [4] for texture classification. Basic LBP operator is a computational efficient operator. Taking each pixel as a threshold, the operator transferred its 3 3neighborhood into a 8-bit binary code, as shown in Figure 5. () Figure 5. Basic LBP operator The decimal form of the resulting 8-bit word (LBP code) can be expressed as follows: 7 i i c (4) LBP= B j -j where ji corresponds to the grey value of the center pixel (x c, y c ), j c corresponds to the grey values of the 8 surrounding pixels, and function B(x) is defined as: i=0 1 x 0 Bx= 0 x<0 (5) Due to its texture discriminative property and its very low computational cost, LBP is becoming very popular in pattern recognition Gabor Filtering Gabor filtering [3] is a wavelet-based method for texture description and classification. Gabor filters extract local orientation and scale information of the texture. Two-dimensional Gabor function can be expressed as: 1 1 x y g(x,y)=( )exp jw x y x y x (6)

5 where w is the frequency of bandwidth, x and y are gaussian constants along x axis and y axis direction. By scale and orientation transformation based on mother wavelet g(x,y), you can get a group of auto-correlation filter: mn -m ' ' g x,y =k g x,y, k 1, m, n Z (7) ' m ' m where x k xcos ysin, y k xcos ysin, indicates the direction, k m has related to scale. For a given image I( xy,, ) its Gabor wavelet can be defined as: mn *, (, ), (8) mn T x y I x y g x x y y dxdy * denotes conjugate complex Difference Matrix Differential matrix [3] is expressed as gray change of image pixels in their neighborhood. A texture element of 3 3 neighborhood of pixels is shown in Figure 6. I a, b a, b x 1 a x 1, y 1 b y 1 represents gray value at the coordinates ab, of image. Given the following definitions: Figure 6. Texture element 1 3 4, 1, x+1, y,, -1 x, y+1, +1, -1 x-1, y+1, A x y I x y I A x y I x y I A x y I x y I A x, y I x 1, y-1 I x+1 y+1 (9) then, A xy, A xy, A xy, 1,, 3, A xy represent the gray change of pixel I (x, y) 4, in the horizontal direction (0 ), vertical (90 ), the main diagonal direction (45 ) and vice-diagonal direction (135 ), respectively. A 1, A, A 3, A 4 are four matrix of the differential matrix of the four directions, respectively.

6 .4. Texture feature fusion Because of the shortcomings of the texture descriptor itself, complexity and uncertainty of Realistic scenes, detection results are often not very ideal when using a single texture features for image splicing detection. Therefore, it is necessary to fuse a variety of texture features description together in the local regional authenticity of blind image detection. On one hand, since the characteristics of different textures in different scenes have different performance, the fusion of different texture features helps to widen the scope of application of the algorithm, and improve the accuracy of detection. On the other hand, when using different texture features in the image splicing detection, if a texture feature is missing, you can set weight of this texture zero, other features will be used to obtain the correct result. This can improve not only the detection accuracy but also the robustness and the adaptability of the algorithm. The texture feature fusion process is shown in figure 7. First, because some texture feature descriptions may have great dimensions, principal component analysis for dimensionality reduction of texture features is introduced in order to reduce computational complexity. Second, because texture characteristic value ranges are very different, features are normalized in order to avoid large similarity deviations caused by direct calculation of characteristic vectors. Third, the Euclidean distance is used to measure the similarity of characteristics: d ( S, N) ( V S V ) (10) where V S corresponds to feature vector of shadow area, and V N corresponds to feature vector of non-shadow area. Finally, the linear combination of several Euclidean distances as criterion functions for authenticity of images is shown in formula (11). D w1d 1 wd w3d 3 w4d 4 (11) w w w w 1 1 where d 1, d, d 3, d 4 represent the Euclidean distances of GLCM features, LBP features, differential matrix features, and Gabor filter features, respectively, and represent the weight of the texture features, respectively. Shadow Area/Lit Area 3 4 N GLCM features LBP Feature differential matrix features gabor filter features PCA PCA PCA PCA Normalization Normalization Normalization Normalization Euclidean Distance d1 Euclidean Distance d Euclidean Distance d3 Euclidean Distance d4 Feature Fusion Result output Figure 7. The fusion process

7 3. Results and discussion Some experiment images are selected from CASIA tampered image detection evaluation database [5], and others are collected by authors. All experiment images are saved in JPEG format. The shadow area and lit area of image are extracted by hand. Four texture features are computed, which are GLCM, Gabor filtering with its center frequency set to 0. and its range of the direction of 0,, LBP local binary patterns, and differential matrix. In order to improve the detection rate and reduce the complexity of the algorithm, principal component analysis (PCA) is applied to reduce the dimensionality of texture features. Features with more than 85% of the contribution to image are selected as main components to achieve balance of efficiency and performance. Initial weights are given according to the experience. Last weights after the experimental results analysis are adjusted as follows: GLCM feature s weight 1 = 0.4, LBP feature s weight = 0.3, the weight of differential matrix 3 = 0., Gabor filtering s weight 4 = 0.1. Value Feature Table 1. The detection results of authentic scene Gray level cooccurrence binary matrix Local Difference Gabor filter matrix Pattern Feature fusion result Figure 1 grass road grass land playground Value Feature Table. The detection results of splicing scene Gray level cooccurrence matrix Local binary Pattern Difference matrix Gabor filter Feature fusion result Figure 6 grass grass land road load Images shown in Figure 8 for testing scenarios include different categories, such as grass, pavement and other different types of scenes. Figure 8(a-d) are real images and Figure 8(e-h) are tamper images. Experiment results are shown in table 1 and table. From the experimental data, we can found that some texture features such as Gabor features are insufficient to distinguish between real image and tamper image. In order to enhance the ability of texture descriptors to distinguish between real image and tamper image and improve the accuracy of each experiment, we made a fusion of different texture features. From experiment results shown in table1 and table, we know that the gray level cooccurrence matrix has the strongest ability to distinguish between different textures. So GLCM is assigned as a larger weight 1 = 0.4, whereas the LBP features and differential matrix feature are set to relatively small weight = 0.3 and 3 = 0., and the weight of Gabor filter features is set as 4 = 0.1. The Euclidean distances of real images shown in Figure 8(a-d) were , , , and , while the Euclidean distance of tamper images shown in Figure 8(e-h) were 1.507, , , and In other words, detection results are more obvious after applying multiple feature fusion. The Euclidean distance of real scenario is maintained above 1.0, while the splitting scene is generally low Euclidean distance values, 0.7 or less.

8 (a)grass (b)grass (c)road (d)playground (e)grass (f)grass 4. Conclusion (g) land (h)land Figure 8. Part of the experiment images where N represents the non-shadow area and S represents the shadow area This paper first introduces the concept of texture consistency of shadow, then proposes a detailed algorithm based on texture consistency of shadow for image splicing detection. Several texture features are fused to improve performance. Experimental results show that the algorithm is simple and effective. However, there are some shortcomings in the experiment and disadvantages, such as manually selecting the shadow and non-shadow extraction areas and setting fusion weights set according to experience. In the future, the authors will continue to optimize the algorithm and further improve the applicability and intelligent process of the algorithm. 5. References [1] Hany Farid, "Detecting digital forgeries using bispectral analysis", Massachusetts Institute of Technology (technical report), 1999 [] Tian-Tsong Ng, Shih-Fu Chang, "Blind detection of photomontage using higher order statistics", In Proceeding of the 004 IEEE International Symposium on Circuits and Systems,pp ,004 [3] Dongdong Fu, Yun Q. Shi, Wei Su, "Detection of image splicing based on Hilbert-Huang transform and moments of characteristic functions with wavelet decomposition", In Proceeding of the 5th International Workshop on Digital Watermarking, pp ,006 [4] Xiaobing Kang, Guangfeng Lin, Yajun Chen, Erhu Zhang, Ganglong Duan, "Detecting Tampered Regions in Digital Images Using Discrete Cosine Transform and Singular Value Decomposition", International Journal of Digital Content Technology and its Applications, vol.6, no.3, 01 [5] Juan Qin, Feng Li, Lingyun Xiang, Jianming Zhang, "Detection of Image Region Copy-move Forgery Using Wavelet Moment", International Journal of Digital Content Technology and its

9 Applications, vol.6, no.18, pp , 01 [6] Yagiz Sutcu, Baris Coskun, Husrev T. Sencar, Nasir Memon, "Tamper detection based on regularity of wavelet transform coefficients", In Proceeding of the 14th IEEE International Conference on Image Processing, pp ,006 [7] Wen Chen, Yun Q. Shi, Wei Su, "Image splicing detection using -D phase congruency and statistical moments of characteristic function", In Proceeding of the SPIE Security, Steganography, and Watermarking of Multimedia Contents IX,007 [8] Hui fan, Yongliang Wang, Jinjiang Li, "Exposing Image Fuzzy Forgeries based on Dyadic Contrast Contourlet", Journal of Convergence Information Technology, vol.6, no.7, pp.38-47, 011 [9] Micah K. Johnson, Hany Farid, "Exposing Digital Forgeries in Complex Lighting Environments", IEEE Transactions on Information Forensics and Security, vol., no.3, pp , 007 [10] Micah K. Johnson, Hany Farid, "Exposing Digital Forgeries by Detecting Inconsistencies in Lighting", In Proceeding of the 7th workshop on Multimedia and security,pp1-9,005 [11] Wei Wang, Jing Dong, Tieniu Tan, "Effective image splicing detection based on image chroma", In Proceeding of the 009 IEEE International Conference on Image Processing, pp ,009 [1] Micah K. Johnson, Hany Farid, "Exposing Digital Forgeries Through Chromatic Aberration", In Proceeding of the ACM Multimedia and Security Workshop,pp.48-44,006 [13] Mo Chen, Jessica Fridrich, Jan Lukáš, Miroslav Goljan, "Imaging sensor noise as digital X-ray for revealing forgeries", In Proceeding of the 9th International Information Hiding Workshop, pp.34-58, 008 [14] Popescu, A.C., Hany Farid, "Exposing digital forgeries in color filter array interpolated images", IEEE Transactions on Signal Processing, Institute of Electrical and Electronics Engineers Inc.,vol.53, no.10, pp , 005 [15] Ashwin Swaminathan, Min Wu, K. J. Ray Liu, "Optimization of input pattern for semi nonintrusive component forensics of digital cameras", In Proceeding of the 007 IEEE International Conference on Acoustics, Speech and Signal Processing, pp.5-8,007 [16] Bo Wang, Xiangwei Kong, Lanying Wu, "Different-quality Re-demosaicing in Digital Image Forensics", Journal of Convergence Information Technology, vol.7, no.17, pp , 01 [17] Yu-Feng Hsu, Shih-Fu Chang, "Image splicing detection using camera response function consistency and automatic segmentation", In Proceeding of the IEEE International Conference on Multimedia and Expo, pp.8-31,007 [18] Yu-Feng Hsu, Shih-Fu Chang, "Detecting image splicing using geometry invariants and camera characteristics consistency", In Proceeding of the 006 IEEE International Conference on Multimedia and Expo,pp ,006 [19] Wei Zhang, Xiaochun Cao, Jiawan Zhang, Jigui Zhu, Ping Wang, "Detecting Photographic Composites Using Shadows", In Proceeding of the ICME 009, pp , 009. [0] Qiguang Liu, Xiaochun Cao, Chao Deng, Xiaojie Guo, "Identifying Image Composites Through Shadow Matte Consistency", IEEE Transactions on Information Forensics and Security, vol.6, no.3, pp , 011 [1] Ning Wang, "Study on Shadow Detection and Removal" (in Chinese), Beijing Jiao Tong University (master dissertation), China, 008. [] Lin Yang, "Research on the Image Segmentation and Shadow Removing Algorithm" (in Chinese), Harbin Institute of Technology (master dissertation), China, 007 [3] Mark Nixon, Alberto Aguado, Feature Extraction and Image Processing (second edition), Publishing House of Electronics Industry, China, 010. [4] Timo Ojala, Matti Pietikäinen1, David Harwood, "A comparative study of texture measures with classification based on feature distributions", Pattern Recognition, Pergamon Press Inc, vol.9, no.1, pp.51-59, 1996 [5] CASIA Tampered Image Detection Evaluation Database,