SURF-based Detection of Copy-Move Forgery in Flat Region

Transcription

1 SURF-based Detection of Copy-Move Forgery in Flat Region 1 Guang-qun Zhang, *2 Hang-jun Wang 1,First Author,* 2Corresponding Author School of Information Engineering, Zhejiang A&F University, , Lin an, China 2. Department of Automation, University of Science and Technology of China, , Hefei, China 1 gloria@zafu.edu.cn, * 2 hangjunw@ .ustc.edu.cn Abstract Techniques for digital image tampering are becoming widespread for the availability of low cost technology in which the image could be easily manipulated. Copy-move forgery is one of the tampering techniques that are frequently used and has recently received significant attention. But the existing methods, including block-matching and key point matching based methods, are not able to be used to solve the problem of detecting image forgery in both flat region and non-flat region. In this paper, combining the thinking of these two types of methods, we develop a SURF-based method to tackle this problem. In addition to the determination of forgeries in non-flat region through key point features, our method can be used to detect flat region in images in an effective way, and extract FMT features after blocking the region. By using matching algorithms of similar blocked images, image forgeries in flat region can be determined, which results in the completing of the entire image tamper detection. Experimental results are presented to demonstrate the effectiveness of the proposed method. Keywords: Speeded-up robust features, Flat-region detection, Image forensics detection, Copy-move forgery 1. Introduction With the widespread use of digital cameras, powerful personal computers, and sophisticated image processing, the manipulation of digital images is becoming much easier. Recently, there have been numerous examples of tampered images in newspapers and on magazine with copy-move forgery technology, where a part of the image itself is copied and pasted into another part of the same image. Figure 1 shows Harmonious coexistence between man and nature in Nanning appeared People's Daily in After the publication of this photo, it was revealed that six Pigeons were digitally added to the image. Therefore, developing techniques to verify the integrity and the authenticity of the digital images became very important. To accomplish this task, many copy-move forgery detection techniques have been proposed in recent years. These methods can be divided in two categories, block-matching and key point matching. Fridrich[1] firstly proposed a method for detecting the copy-move forgery, which apply discrete cosine transform (DCT) to the block. Popescu [2] applied PCA on small fixed-size image to yield a reduced-dimension DCT block representation. Duplicated regions are then detected by lexicographically sorting and grouping all of the image blocks. Li [3] used the features Singular Value Decomposition (SVD) performed on low-frequency coefficients of the block-based Discrete Wavelet Transform (DWT). Luo [4] chose the features of red, green and blue components together with other four computed on the blocks by calculating the energy distribution of luminance along four directions. Mahdian and Dybala focused on the detection of replicated regions affected by blurring and filtering, respectively [5, 6]. Yanjun Cao[7]used fast Pseudo-Zernike moment. Yanjun Cao[8] used less features to represent each block. International Journal of Advancements in Computing Technology(IJACT) Volume4, Number17,September doi: /ijact.vol4.issue

2 Figure 1. An example of copy-move forgery The above methods are most effective, but are not able to detect scaling or rotation transformations. So the new research emerged to tackle these problems. Bayram [9] proposed the Fourier Mellin Trans-form (FMT) on the image blocks, which is robust to special case of compression, rotation and scaling. Hwei-Jen Lin [10] proposed a method which performed well when the degree of rotation was 90, 180 and 270 degree. Bravo-Solorio [11, 12] proposed reflection, rotation and scaling -invariant features by log-polar coordinates mapping. Seung_Jin Ryu [13] used Zernike moments to obtain rotation-invariant features. Wang [14] used the mean intensities of circles with different radii around the block center. Comparison among the copy-move methods discussed above has been reported in [15-17] for evaluating their performance. All the above methods are block-matching based, which divide the image into overlapping blocks and extract features from each of them. A matching step then takes place to find the duplicated blocks. But most of the existing blocking matching methods are not good in detecting the image forgery with manipulations including loss compression, Gaussian noise, rotation, scaling, and also are time-consuming. So another category algorithm, key point matching based methods are proposed. Huang and Amerini extracted SIFT feature of an image, and matched between each other to seek for any possible forgery [18, 19]. Pan also extracted SIFT descriptors to detect duplicated regions with continuous rotation regions, which is effective to synthesized forgery image with duplicated and distorted regions [20, 21]. Although Using key point matching based methods guarantee geometric invariance, there is no key point on smooth background, which means it may fail to detect object concealing forgery. Smooth background also call flat region, which is defined as such kind of region with low frequency and/or smooth gradient. In flat region, pixel values are similar with each other and varying smoothly over relatively large region. So in the paper, we propose an efficient method for detecting copy-move forgeries in digital images, combining the advantages of these two types of methods, which divide an image into two parts, namely, non-flat region and flat region. Non-flat region use the key point matching based method, and flat region use block matching. The rest of this paper is organized as follows. Related works on extracting key point features with SURF and FMT features from image blocks is discussed in Section 2. In Section 3, the proposed method is described in details. In Section 4 and 5, we show some experimental results and make a conclusion for this paper, respectively. 2. Related works 2.1. Speeded-Up Robust Features extraction As a local feature descriptor, SURF [22, 23] is introduced here because it has a discriminative power, which is robust and can be made invariant to rotation. It is similar to SIFT features, which have been successfully used for image forgery detection [18-21], but can be computed faster. SURF algorithms can each be divided into two distinct parts: interest point detection, which attempts to find stable interest points in the image, and interest point description, which generates descriptors for every interest point found in the detection step. For the extraction of SURF features, in this paper, we use the OpenSURF to detect image tempering in non-flat region for its efficiency in computing, and the detailed overview and implementation is provided in [24]. 522

3 2.2. Fourier-Mellin Transform based features extraction Recently, a new efficient block-matching based copy-move forgery detection method referencing to the Fourier-Mellin Transform (FMT) was proposed by Bayram [9], which performs radial projection on the log-polar coordinate Fourier transformation of image blocks. For a block i(x,y), the following procedure is applied to this blocks: 1) Obtain the Fourier transform representation to ensure features are translation invariant x y x y x y I'( f, f ) I'( ( f cos, f sin ), ( f sin, f cos )) (1) 2) Re-sample the resulting magnitude values into log-polar coordinates 2 I'(, ) I( log ), (2) 3) Project log-polar values onto 1-D, and obtain 45 features by quantizing these summed values for different. g( ) log( I( j, ) j (3) FMT features showed a very good overall performance; therefore they are used in this paper for forgery detection of flat region. 3. Proposed method We think that an image consists of two parts, flat and non-flat region. Taking into account the efficiency and accuracy of forgery detection algorithm, we propose the method in which these two regions are firstly separated by using flat region detection algorithm. We then detect the forgery regions with block-matching based method and key point matching based method respectively. A brief procedure of our method is shown in Table 1. Table 1. Brief procedure of our method 1 Forgery detection in non-flat region 1.1 Key points detecting using SURF. 1.2 Feature matching and pruning. 1.3 Estimating region transforms and identifying duplicated regions using correlations adjusted with the estimated transforms. 2 Flat region detection 2.1 Draw square taking key points as the center. 2.2 Erosion and Open operators in mathematical morphology are used to fill small holes. 2.3 Fill the small closed regions. 3 Image blocking and feature extraction 3.1 Image blocking in flat region. 3.2 FMT feature extracting from the image blocks. 4 Identifying duplicated regions 4.1 Radix sort to find similar blocks 4.2 Forgery decision 4.3 edge processing of detected forgery region 3.1 Forgery detection in non-flat region First, by using SURF discussed in Section 2.1, we extract key points from image. These points are exploited to reconstruct the parameters of the occurred geometric transformation, which is based on the Xueyun Pan s frame [20], consisting of three steps: (i) Key points detecting using SURF, (ii) Feature 523

4 matching and pruning, and (iii) Estimating region transforms and identifying duplicated regions using correlations adjusted with the estimated transforms. It can be used to detect duplicated region, unlike the recent work [19] using SIFT key point matching to estimate the parameters of the affine transform and recover matched key-points, which only displays the matched key points. 3.2 Flat region detection In flat region, since pixel values are similar with each other and varying smoothly over relatively large region, we can detect these regions by observing the changes of local pixel values. Ahn [25] present a cost-efficient method to detect flat region in the image, using the entropy and the second-order statistics, local deviation of the local region, centered at the current pixel. But the purpose of flat region detection in this paper is to locate the regions that key point matching based methods fails to handle. So we need a better algorithm to deal with this problem. Our proposed method to detect the image flat region is based on key points obtained in Section 3.1. Assuming the n key points in the image are p1, p2,, pn., the flat region map, FLM, is generated as shown in Table 2. After running the generation algorithm, we get the white area, presented in FLM, which is exactly the flat region. Table 2. FLM generation algorithm Step 1 Initializing: FLM(i,j)=255 (denote as white) for all pixels in image. Step 2 For the n key points, pi, draw the black square taking the key point as the center and length a. An example is shown in Figure 2 (b), where the key points are shown in red. Step 3 Erosion and Open operators in mathematical morphology are then used to fill small holes in the image. As a result, we get the image shown in Figure 2 (c). Step 4 Finally, we fill the closed regions with area less than A, and then get Figure 2 (d). (a) key points (b) draw squares (c) Erosion and Open operators (d) fill small closed region Figure 2. An example of the procedure of flat region detection. 3.3 Image blocking and feature extraction Through the flat region detection algorithm, in the flat region map, FLM, with the grey value at 255, white part is the flat region. As the region is composed of a number of irregular shapes, our method blocks the flat region with fixed-size rectangular. 524

5 Firstly, blocking algorithm divide FLM of the image into fixed-size non-overlapping blocks from left to right, and then from top to bottom. Delete these sub-blocks, where there is no one point belonging to the flat region. Figure 3 shows an example of image blocking. Such way of image blocking brings two benefits: 1) Neglecting non-flat region; 2) reduce the number of blocks in the flat region. But it leads to a problem that: how to detect the tampering region within part of a block? In Section 3.4, we will solve this problem by edge operation for tampering region. Figure 3. An example of image blocking on flat region FMT then is applied to each block to represent its features with FMT based feature extraction method discussed in Section 2.2. Firstly, take the Fourier transform to ensure features are translation invariant, and extract frequency magnitude along directions as features, whereas frequency magnitude along log( ) directions are projected together. We compute for from degree 0 to 180 with step 2; then add two halves together. So, at last 45 features quantized to integer are obtained. 3.4 Identifying duplicated regions For finding similar blocks, we use radix sort to sort the feature vectors of the divided blocks, because it improves the computational complexity in detecting the regions of forgeries, as an alternative to lexicographic sorting, which is commonly used by the existing copy-move forgery detection schemes [10]. After obtaining feature vectors for all blocks as described in Section 3.3, matrix A is constructed with these feature vectors in a way that the rows of A correspond to the blocks and columns indicating the feature vectors. For the block size b*b, and if there are totally k blocks, the number would be far smaller than the number of block used in the state-in-art methods. Through sorting the matrix A by radix sort, the difference of the positions of every pair of adjacent feature vectors in the sorting list, defined as shift vector, is computed. The accumulated number of each of the shift vectors is evaluated. A large accumulated number u is considered as possible presence of a duplicated region, and thus all the feature vectors corresponding to the shift vectors with large accumulated numbers are detected, whose corresponding blocks are then marked to form the detected result. Owing to using the non-overlapping blocking way, when obtaining the detected result, we need a technology, called edge processing, to refine the forgery region detected. Figure 4 shows an example of the forgery region detected. We assume the region enclosed by the red line is the real tampered region; the region enclosed by the black thick polygon (consisting of B, internal edge of the region, and C, internal region) is the detected region obtained through the above detection algorithm. Table 3 gives the detail algorithm of edge processing [26]. 525

6 Figure 4. An example of image tampering Table 3. A algorithm for edge processing of the detected forgery region Step 1 Find the internal edge of the detected tampered region, which consists of image blocks where there is a block not belonging to the detected tampered region in their 4 neighborhood image blocks. Step 2 Find the external edge of the detected tampered region, which consists of the 8 neighborhood image blocks of the internal edge obtained in the Step 1, which don t belong to the detected tampered region. Step 3 Block all the external edge in a smaller size. An example is shown in the left-top part of Fig. 4. Then for these smaller image blocks, feature extraction and duplicate region detection is also carried out according to the method in Section 3.3 and 3.4 to obtain a more accurate tampered region. 4. Experimental results and analysis In this section, we describe experimental evaluations of the proposed forgery detection methods. Images used in the experiments are from our forgery image database, which is composed of 300 images: 150 are tampered images and 150 are original images. The images resolution varies from 500*375 to 800*600 pixels. Firstly, we choose the images from it, where flat region and non-flat region are both tampered. Because the aim of flat region tamper in general is to hide unwanted parts of the image, flat region neighboring was copied to hide them. It makes them (the parts don t wanted be show in the image) also become flat region. Usually, the flat region tamper seldom uses image transformations. But non-flat region of the images is tampered by applying diverse transformations. So we test over 10 tamped images with compression or Gaussian on flat region, and with rotation or scaling on non-flat region chosen from above images. The two detected results are shown in Fig. 5 and 6. Fig. 5 shows the image detected results including tempered region with scaling and Fig. 6 with rotation. The proposed approach has been compared with the implementations of the method presented in [18]. The experiments are implemented on a computer of CPU 2.60GHz with Memory 2GB. The input parameters required by the method are set as follows: All the sub-block size use b=16 except 4 in the edge processing for detecting forgery region; The OpenSURF parameters are octaves=4, intervals=1, init_sample=2, thres= f, and the correlation threshold for contour blocks is 0.3, and the area threshold is 600 pixels in non-flat region s forgery detection; square length a=16, the threshold of filling closed regions area A=500 in Flat region detection as we think that the tempered region has a certain size, then can be noticed; u=6 (the accumulative number of a shift vector). In Fig. 5(b), there are two positions tempered: the roof in the left of the image has been erased and a copied tower in the right appears in the vicinity. The method based on key point matching firstly extracts key points from the image, which are shown in Figure 5(c). The regions of the two towers will form the same image area for their similar key point features, so they can be detected, and shown in Fig. 5(d). But the roof can not be detected because there is no key point in the forgery regions. Based on our approach, in addition to the detection of forgery regions using the key point based methods, we also generate the flat region, which is shown in Figure 5(e). It is blocked and then extracted FMT features. After image block matching, an accurate image forgery result is shown in Figure 5(f). 526

7 (a)original image (b)tampered image (c)key points extraction (d)detection result with key point based method (e)the result of flat region detection (f) detection result with our method Figure 5. An example of detecting copy-move forgery For the two image tampered regions in Figure 6(b), one is the upper part of the street lamp located in the right is covered with the background, and the other is that the cycling man on the right car is copied, and placed in the left rear of the car after a certain right-handed angle processing. In like manner, the result using key point matching based detection method is shown in Figure 6(c), where two cycling men with the same key points were detected. While it fails to detect the street lamp because there is no key point in it. Based on our method, the result is shown in Figure 6(d), where all the tempered image regions including the street lamp and the cycling man have been detected. (a)original image (b)tampered image (c) result with key points (d) result with our method Figure 6. Another example of detecting copy-move forgery 527

8 These two image example results point out that the proposed method performs better with respect to the other method. Blocking matching based methods are able to detect flat region tamper, but not able to detect non-flat region tamper with certain types of transformations in an effective manner. Also, the processing time (per image) is on average about 20 seconds on our selected images. Whereas our method can detect all tamper regions, and need the relatively smaller run time, about 5 second on average. 5. Conclusions Forgery images created with duplicated and distorted regions are challenges to detect visually. In this paper, a new forensic method has been proposed to detect and localize duplicated regions in both flat regions and non-flat regions. This problem has not been addressed in current literature. Because key point matching based methods will be powerless when meeting the no key points regions, but where may also exist tempering. So, by separating the two regions of the image, different methods were used to detect in our approach, which achieve maximum efficiency and effectiveness of the forgery detection. Having achieved promising performance in detecting sophisticated forgeries existing in both flat region and non-flat region, our method relies on the detection of flat region. For large number of images auto-detection, this may be a limitation as many parameters need to be automatically determined, such as the length of square, the size of structure element, and the area of closed region, etc. So it s a main challenge for the proposed method. As one of the important future works, we will consider several approaches to improve the performance of detection flat region, and the two forgery detection methods. 6. Acknowledgment This work was supported by Educational Commission of Zhejiang Province of China, No. Y and by Zhejiang A&F University research project, No. 2010FK References [1] Fridrich J., Soukal D., Lukáš J., Detection of Copy-Move Forgery in Digital Images, in Proceedings of Digital Forensic Research Workshop, pp.19-23, [2] Popescu A., Farid H, Exposing digital forgeries by detecting duplicated image regions, Technical Report, Dept. Comput. Sci, Dartmouth College, Tech. Rep. TR , pp.1-11, [3] Li G., Wu Q., A sorted neighborhood approach for detecting duplicated regions in image forgeries based on DWT and SVD, Proc. IEEE ICME, pp , [4] Luo W., Huang J., Robust detection of region-duplication forgery in digital image, In Proceedings of the 18th International Conference on Pattern Recognition, pp , [5] Mahdian B., Saic S, Detection of copy-move forgery using a method based on blur moment invariants, Forensic Science International, vol. 171, no. 4, pp , [6] Dybala B., Jennings B., Letscher D, Detecting filtered cloning in digital images, Proc. MM&SEC Workshop on Multimedia and Security, pp.43-50, [7] Yanjun Cao, Tiegang Gao, Li Fan, Qunting Yang, "A Robust Detection Algorithm for Region Duplication in Digital Images", JDCTA, Vol. 5, No. 6, pp , 2011 [8] Yanjun Cao, Tiegang Gao, Li Fan, Qunting Yang, "A Novel Approach for Detecting Region Duplication in Digital Images", AISS, Vol. 3, No. 10, pp , 2011 [9] Bayram S., Sencar H.T.,Memon N, An effcient and robust method for detecting copy-move forgery, Proc. IEEE CASSP, pp , [10] Lin H-J., Wang C-W., Kao Y-T, Fast Copy-Move Forgery Detection, WSEAS Transaction on Signal Processing, vol.5, no. 5, pp ,2009. [11] Bravo-Solorio S., Nandi AK, Passive forensic method for detecting duplicated regions affected by reflection, rotation and scaling, Proc.European Signal Processing Conference, pp ,

9 [12] Bravo-Solorio S., Nandi AK, Automated detection and localisation of duplicated regions affected by reflection, rotation and scaling in image forensics, Signal Processing, vol.91,no.8, pp ,2011. [13] Ryu S-J., Lee M-J.,Lee H-K, Detection of Copy-Rotate-Move Forgery using Zernike Moments,Proc.12th International Workshop on Information Hiding, pp.51-65,2010. [14] Wang J., Liu G., Li H., Dai Y., Wang Z., Detection of Image Region Duplication Forgery Using Model with Circle Block, Proc.2009 International Conference on Multimedia Information Networking and Security, pp.25 29,2009. [15] Christlein V., Riess C., Angelopoulou E., A study on features for the detection of copy-move forgeries, Proc. GI-Edition Lecture Notes in Informatics, pp , [16] Christlein V., Riess C., Angelopoulou E., On rotation invariance in copy-move forgery detection, Proc. the 2010 Second IEEE Workshop on Information Forensics and Security, pp.1-6, [17] Shivakumar B.L., Baboo S.S, Detecting Copy-Move Forgery in Digital Images: A Survey and Analysis of Current Methods, Global Journal of Computer Science and Technology, vol.10, no.7, pp.61-65, [18] Huang H., Guo W., Zhang Y, Detection of Copy-Move Forgery in Digital Images Using SIFT Algorithm, Proc. IEEE Pacific-Asia Workshop on Computational Intelligence and Industrial Application, pp , [19] Amerini I., Ballan L., Caldelli R., Bimbo A.D., Serra G., Geometric Tampering Estimation by Means of a SIFT-based Forensic Analysis, Proc. IEEE International Conference on Acoustics Speech and Signal Processing (ICASSP), pp , [20] Pan Xunyu., Lyu Siwei., Detecting Image Region Duplication Using SIFT Features, Proc. IEEE International Conference on Acoustics Speech and Signal Processing (ICASSP), pp , [21] Pan Xunyu., Lyu Siwei., Region Duplication Detection Using Image Feature Matching, IEEE Transactions on Information Forensics and Security, vol.5,no.4, pp ,2010. [22] Bay H., Tuytelaars T., Van Gool LV, Surf: speeded-up robust features, Proc. 9th European Conference on Computer Vision, pp , [23] Bay H., Ess A., Tuytelaars,T.,Gool LV, Speeded-up robust features(surf), Computer Vision and Image Understanding (CVIU), vol.110, no.3, pp , [24] Christopher Evans., Notes on the opensurf library, University of Bristol, USA, Technical Report CSTR , [25] Ahn W., Kim JS, Flat-Region Detection and False Contour Removal in the Digital TV Display, Proc. ICME, Amsterdam, Netherlands, pp , [26] ZHANG G-q., WANG H-j, Quick detection of Copy-Move forgery based on FMT, Computer Engineering and Design, vol.31, no.15, pp ,