Keypoints based enhanced multiple copy‐move forgeries detection system using density‐based spatial clustering of application with noise clustering algorithm
2018; Institution of Engineering and Technology; Volume: 12; Issue: 11 Linguagem: Inglês
10.1049/iet-ipr.2018.5576
ISSN1751-9667
AutoresBadal Soni, Pradip K. Das, Dalton Meitei Thounaojam,
Tópico(s)Advanced Steganography and Watermarking Techniques
ResumoIET Image ProcessingVolume 12, Issue 11 p. 2092-2099 Research ArticleFree Access Keypoints based enhanced multiple copy-move forgeries detection system using density-based spatial clustering of application with noise clustering algorithm Badal Soni, Corresponding Author Badal Soni soni.badal88@gmail.com Computer Science & Engineering, National Institute of Technology Silchar, Assam, IndiaSearch for more papers by this authorPradip K. Das, Pradip K. Das Computer Science and Engineering, Indian Institute of Technology Guwahati, Assam, IndiaSearch for more papers by this authorDalton Meitei Thounaojam, Dalton Meitei Thounaojam Computer Science & Engineering, National Institute of Technology Silchar, Assam, IndiaSearch for more papers by this author Badal Soni, Corresponding Author Badal Soni soni.badal88@gmail.com Computer Science & Engineering, National Institute of Technology Silchar, Assam, IndiaSearch for more papers by this authorPradip K. Das, Pradip K. Das Computer Science and Engineering, Indian Institute of Technology Guwahati, Assam, IndiaSearch for more papers by this authorDalton Meitei Thounaojam, Dalton Meitei Thounaojam Computer Science & Engineering, National Institute of Technology Silchar, Assam, IndiaSearch for more papers by this author First published: 05 September 2018 https://doi.org/10.1049/iet-ipr.2018.5576Citations: 11AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract In this study, the problem of detecting if an image has tampered is inquired; especially, the attention has been paid to the case in which the portion of an image is copied and then pasted onto another region to create a duplication or to hide some important portion of the image. The proposed copy-move forgery detection system is based on the scale-invariant feature transform (SIFT) features extraction and density-based clustering algorithm. The extracted SIFT features are matched using the generalised two nearest neighbours (2NN) procedure. Thereafter, the density-based clustering algorithm is utilised to improve the detection results. The proposed system is tested using MICC-F220, MICC-F2000 and MICC-F8multi datasets. Due to the generalised 2NN matching procedure, the proposed system is able to detect multiple forgeries present in the image. Experimental results show that the performance of the system is quite satisfactory in terms of computational time as well as detection accuracy. 1 Introduction Copy-move forgery is a popular image manipulation technique, in which, the attacker creates the multiple instances of some object(s) or sub-portion(s) of the image by copying and pasting it into different locations of the same image. This procedure generates an image which is nothing but a copy-moved forged image. Since the copied portion comes from the same image, the colour quality, noise components, intensity range and other image properties will be almost the same with the rest of the image, therefore it is difficult to detect copy-move forgery in images. The task of any copy-move forgery detection (CMFD) system is to determine the portion(s) of the image, which is exactly matched to some other portion(s) of the same image. Sometimes to make the copy-move forgery unnoticeable to the user, some modifications such as rotation, scaling, noise embedding, filtering etc. are performed either on the whole tampered image or only the copied region of the image before pasting it on the original image. CMFD in images is necessary since, in many applications and occasions, images are used as the primary element of communication for sharing the useful information. For an example, in law and order places, images must be authentic and it is necessary to ensure its authenticity. In these types of places, forgery detection in images becomes a key requirement. Therefore, detection of copy-move forgery has become a prominent research area. Depending on the processing ways, the CMFD may be classified into two broad categories, i.e. keypoints and block-based techniques. The block-based techniques are generally started by partitioning the image into either overlapping or non-overlapping blocks. Thereafter, block features are extracted and matched for forgery detection. Since in large-sized images, the number of blocks is more, thereby the computational cost of the block-based methods is high. In addition to that, most of the block-based approaches are sensitive against geometric transformations, i.e. rotation, scaling and reflection. Moreover, keypoints-based approaches, to some extent reliably compensate for these drawbacks. In these approaches, local image features such as corner points, blobs and edges are extracted from the input-forged image. The features are represented by a set of descriptors to increase the reliability of the features and matching is performed among the feature descriptors to find the forged regions in the image. Recently, a detailed survey on keypoints and block-based copy-move forgery techniques with a critical discussion of their pros and cons has been given in [1]. In this study, we propose an improved system for CMFD using the scale-invariant feature transform (SIFT) keypoints and density-based spatial clustering of application with noise (DBSCAN) clustering algorithm. Initially, the SIFT keypoints are detected from the image and the corresponding descriptors are extracted and matched using the generalised two nearest neighbours (2NN) matching procedure given in [2]. Thereafter, the density-based clustering approach is utilised to decrease the false matches and to improve the detection accuracy of the system. In the proposed system, multiple forgeries are handled by performing a robust generalised 2NN matching procedure and then by performing DBSCAN clustering on the keypoints spatial position in order to separate the different copied regions. The main contributions of this study are: Propose an improved system, which is based on the SIFT keypoints and density-based clustering algorithm and is able to detect multiple forgeries present in the image. In the proposed system, extracted SIFT features are matched using the generalised 2NN procedure. Thereafter, the density-based clustering algorithm is utilised to improve the detection results. The utilisation of the density-based clustering algorithm leads to the reduction of the number of false matches, thereby reducing the false positive rate (FPR) and ultimately increasing the accuracy of the system. The paper is organised as follows: Section 2 address the existing work in CMFD. Section 3 describes the background knowledge. The proposed methodology is given in Section 4. Experimental results and discussions are detailed in Section 5. Finally, conclusions and future scope are given in Section 6. 2 Related work In the last decade, there has been a lot of research on the detection of copy-move forgery from digital images. The very first work of CMFD was carried out by Fridrich et al. [3]. In this method, four different matching procedures such as exact block match, autocorrelation, exhaustive block search and robust match have used for forgery detection. Among all these, the robust matching procedure has been producing efficient and accurate detection results. However, this procedure leads to many false matches when applied to the large identical textured regions image. In [4], the SIFT keypoints are utilised for CMFD. In this study, the extracted SIFT keypoints are matched using Euclidean distance for forgery decision. In [5], a method to detect multiple copies of forged regions in a tampered image was proposed. This method is divided into three phases: SIFT point clustering, cluster matching and texture analysis. In this work, the keypoints are extracted by SIFT. The decision of forgery is taken in the matching step where the clusters of keypoints are matched to each other, rather than a single keypoint. Object shape along with texture analysis is used to compare the content of the two matching objects. It is observed that the FPR of this method is lower than that in [4]. However, this method under-performs in small smooth images. A generalised 2NN procedure for SIFT descriptors matching was proposed in [2]. The accuracy of this method is quite impressive in the MICC-F220 dataset. However, this technique fails to detect a smooth or uniform texture copied image patch where salient key-points are not detected by SIFT. Using J-Linkage clustering the improvement of [2] is given in [6]. In this method, the localisation of the copied region has been done based on the clusters obtained using the J-Linkage algorithm. In [7], the best bin first (BBF) approach was proposed for matching the SIFT descriptors. After that, the random sample consensus (RANSAC) approach is utilised for eliminating outlier points. It seems that this technique is robust against geometric transformation attacks. However, this system leads to inaccurate results due to the presence of numerous mismatch points. The combination of locality preserving projection (LPP) and SIFT was used in [8]. SIFT featuring dimension is reduced by LPP after those descriptors are matched using the Euclidean distance similarity measure. It is observed that due to the dimension reduction property of the LPP, this technique is fast and performs well in the presence of rotation, scaling and compression attacks. However, this method failed in small uniform regions forgery detection. In [9], speeded up robust features (SURF) descriptors were utilised for CMFD. In this method, the extracted SURF descriptors matched forgery decision. It seems that this technique is fast and efficient for low resolution small images. However, in this technique, forgery localisation is not performed. In [10], the combination of discrete cosine transform (DCT) and SURF was proposed for forgery detection. This technique is tested using the UCID database. It seems that the technique is able to detect and localise the forgery in the input forged images. However, this technique is tested only on one small-sized dataset. Hierarchical agglomerative clustering (HAC) was introduced for forgery detection in [11]. Initially, SURF descriptors are extracted from the image and HAC is used for cluster matching of the SURF descriptors. It is observed that this method is fast due to the HAC matching. However, the detection accuracy of this method is quite low. CMFD-SIFT, a new keypoint distribution strategy was proposed in [12]. It seems that this technique is able to detect forgery in the flat or smooth region and is robust to geometric and mirror transformation. The new iterative procedure for forgery detection was proposed in [13]. In [14], a combination of SIFT and KAZE interest point descriptor was used to extract more keypoints to detect small uniform area forgery. In [3], a block-based method using the quantised DCT coefficients was proposed. This method is robust to noise, compression and retouching. However, it is unable to detect forgery in the presence of geometric transformation attacks. In [15], a histogram of orientated gradient based statistical features is used for forgery detection. It can be observed from the results that the algorithm needs improvement for detection of forgery in the case of rotation and scaling performed over the large region. In [16], a multi-level dense descriptor (MLDD) extraction method and a hierarchical feature matching technique were proposed. The MLDD extract the dense feature descriptors from each pixel and hierarchical feature matching is used to detect forged regions. In [17], an improved block-based technique was proposed by dividing the image into circular blocks and extracting local and inner image features using discrete radial harmonic Fourier moments (DRHFMs). It is observed that the computational cost of this method is more than the state-of-the-art forgery detection methods due to the overlapping circular blocks division and DRHFMs. A hybrid technique was proposed in [18]. In this technique, the image is partitioned into non-overlapping blocks. After that, simple linear iterative clustering (SLIC) is applied and SIFT descriptors are extracted from all blocks. Depending upon the SIFT descriptors the regions are segmented into smooth and non-smooth regions. The reliability and efficiency of this technique depend upon the SIFT and Zernike moments, respectively. In [19], a blur invariant forgery detection method was proposed. In this method, the input image is first divided into overlapping blocks and the extracted fast Walsh Hadamard transform features of each overlapping block are matched for forgery decision. In [20], a new method for image CMFD based on the local binary pattern (LBP) histogram Fourier features was presented. Similarly, in this method, the image is first divided into overlapping blocks and the extracted LBP histogram Fourier features of each block are matched for forgery decision. A comprehensive survey of CMFD approaches was given in [21]. In this study, the pros and cons of the existing CMFD approaches have been highlighted along with future directions to resolve the limitation of the existing techniques. After going through the publications in the literature, it is found that SIFT-based techniques are the best choice for CMFD, because SIFT keypoint extractions are computationally efficient as well as invariant to the geometric transformations. However, it is also found that the performance of the SIFT-based techniques is low in flat areas or highly similar regions since there is no keypoint present in that area. To overcome the limitation of SIFT and detect forgery in the smooth region, researchers have used some other feature descriptors with SIFT which are capable of performing flat regions forgery detection such as CMFD-SIFT, KAZE and Zernike moments. Still, this combination of features needed some improvements in order to detect proper forgery and handling multiple forgeries present in the image. In this study, we proposed an improved system based on the SIFT keypoints and DBSCAN clustering algorithm with the objectives to detect accurately the multiple forgeries present in the tampered image and reduced computation time. 3 Background knowledge In the proposed system, a keypoint-based feature namely, SIFT, the generalised 2NN matching procedure and density-based clustering approach are used. 3.1 Scale invariant feature transform SIFT, introduced by Lowe [22], is the most widely used keypoint feature in image CMFD. Scale-space extrema detection: Let us consider an input image . The scale space of image I is defined as follows: (1)where * is the convolution operator and G is a Gaussian function given by (2)where is the scale space factor. The difference of Gaussian function is convolved with the image I, for detecting scale and orientation invariant SIFT interest points. This is expressed as (3)or it can be expressed as . The convolved images make a group of the octave and the k value is selected in such a way that it will produce a fixed number of blurred images in each octave. Keypoint localisation: A scale-space extrema step produces a lot of keypoints in which some of the keypoints are unstable. So, in this step, unstable keypoints are filtered and only the resultant stable keypoints are retained for further processing. Orientation assignment: To make the keypoints invariant to rotation, in this step one or more orientations assigned to each keypoint based on the local image gradient directions. For determining the keypoint orientation, a gradient orientation histogram is computed in the neighbourhood of the keypoint. For an image sample at the keypoint scale the gradient magnitude and orientation are computed using pixel differences (4) (5)where and . Keypoint descriptors: In previous steps, the keypoints are detected. These keypoints are invariant to location, scale and rotation. In this step, the SIFT descriptors are computed at the keypoints locations in both image plane and scale-space. The histograms contain eight bins each and each descriptor contains a array of 16 histograms around the keypoint. So, each feature descriptor consists of a histogram of elements. 3.2 Descriptor matching After obtaining the SIFT interest points and their corresponding feature descriptors , descriptor matching is performed by comparing each keypoint descriptor with the remaining descriptors. SIFT descriptors are of a 128-dimensional feature vector; therefore, an exhaustive matching approach is not a good option. Instead of exhaustive matching, we can use 2NN a more effective procedure, as suggested in [22]. In the 2NN procedure, the distance ratio between the closest neighbour and second-closest neighbour is compared with a threshold. For a given keypoint, let the distance vector . It represents the sorted Euclidean distances of each descriptor with other descriptors (6)keypoint is matched only if (6) is satisfied. The threshold value is chosen generally. This matching procedure has the drawback, the 2NN obtains only matches between keypoints whose corresponding descriptors are globally distinctive. Therefore, in the proposed system, we utilise a well-known generalisation 2NN procedure proposed by Amerini et al. [2]. This matching procedure is based on iterating the 2NN procedure between until this ratio is greater than . If l is the value at which the procedure terminates, the keypoint which corresponds to the distances (where ) is considered as a match for the inspected keypoint. By repeating this procedure over all the keypoints K, finally, the set of matched points is generated. 3.3 Density-based clustering algorithm DBSCAN [23] is a data clustering algorithm. It can be stated as for a given set of points in any space, DBSCAN group points are packed closely. In others, sense points have many closest neighbours. In the DBSCAN clustering approach, the points are categorised as core points, density reachable points and outliers in the following ways: A point X is a core point if it has more than a specified number of points within the distance , where is the maximum radius of the neighbourhood of X. These points are said to be directly reachable from X. A point Y is reachable from X if there is a path with and , where each is directly reachable from . An outlier point is any point that is not a core point or a density reachable point. To illustrate the DBSCAN algorithm, an example is given in Fig. 1. In this example, . Point P and the other blue solid points are core points since the area covered by these points in distance consist of minimum four points including the point itself. Since all these points are reachable from one another, therefore, they construct a single cluster. Points Q and R are non-core points and are reachable from core point P using other core points hence belong to the cluster as well. Point S is an outlier point that is neither a core point nor directly reachable point. Fig. 1Open in figure viewerPowerPoint Example of DBSCAN algorithm If X is a core point, then it forms a cluster together with all points (core or non-core) that are reachable from it. A cluster C in a set of points S with respect to and is a non-empty subset of S satisfying Maximality: For all X, Y if and if Y is density reachable from X with respect to and , then also . Connectivity: For all , X is density connected to Y with respect to and in S. Algorithm 1 (Fig. 2) consists of the pseudo-code of the DBSCAN clustering approach. Fig. 2Open in figure viewerPowerPoint Algorithm 1: DBSCAN 4 Proposed methodology The proposed system is based on the extraction of SIFT features, which are used to identify whether a portion of the image is copy-moved or not. Since the copied portion has generally the same appearance as the original one, thus keypoints extracted from the forged portion will be similar to the original portion. Therefore, matching of SIFT features can play a vital role in determining forgery present in the images. In the proposed system, we utilised the generalised 2NN matching and DBSCAN clustering approaches. Algorithm 2 (Fig. 3) consists of the pseudo-code of the proposed system. The proposed system is textually analysed as follows. Fig. 3Open in figure viewerPowerPoint Algorithm 2: Proposed system In the first step, the keypoint features are extracted using the SIFT feature extraction technique given by Lowe [22]. The SIFT keypoints are robust to the rotation and scaling geometric transformation as claimed and used by Amerini et al. [2]. A set of SIFT vectors representing the location, scale, and orientation is assigned to each feature point in each image. Successively, the SIFT descriptors are computed at keypoints locations in both image plane and scale-space. The histograms contain eight bins each and each descriptor contains a array of 16 histograms around the keypoint. So, each feature descriptor consists of a histogram of elements. After that, a well-known generalisation 2NN matching procedure [2] is utilised for matching the extracted SIFT descriptors. The generalisation 2NN matching procedure is based on the iterating the 2NN procedure between until this ratio is greater than . A detail description of the matching procedure is given in Section 3.2. Finally, for discovering the possible copied regions, a density-based clustering [23] is applied to the spatial locations of the SIFT matched points. In the DSSCAN clustering algorithm, the matched points are categorised as core, density reachable points and outliers based on the following procedure; if X is a core point, then it forms a cluster together with all points (core or non-core) that are reachable from it. A cluster C in a set of points S with respect to and is a non-empty subset of S satisfying maximality and connectivity. In the proposed system, multiple forgeries are handled by performing a robust generalised 2NN matching procedure and then perform DBSCAN clustering on the keypoints spatial position in order to separate the different copied regions. These are the primary steps in the case of multiple forgeries, otherwise, it is not feasible to detect and mark separately each forgery. 5 Experimental results and discussions In this section, the description of the different benchmarking databases used for the experimental purpose, performance measures, robustness test of the proposed algorithm, analysis of the experimental results for three different datasets and comparison of the proposed approach with the existing approaches is given. 5.1 Database The proposed system is evaluated using different standard publicly available datasets. The performance of the proposed system is evaluated on three datasets. In this work, we report the experimental results on MICC-F220, MICC-F2000 and MICC-F8multi [2] database images. The MICC-F220 dataset images consist of an average 1.2% forged region of the complete image. The MICC-F2000 dataset consists of a total of 2000 images out of which 700 are forged and 1300 are original images. The dataset MICC-F8multi contains eight realistic and challenging multiple forged attack images. The description of the used datasets in the proposed system is given in Table 1. Table 1. Datasets description Dataset Image size No. of images MICC-F220 to 220 MICC-F2000 2048 × 1536 2000 MICC-F8multi 2048 × 1536 8 5.2 Performance measures Experiments are performed using an HP machine, Intel Core i5-3230M (2.60 GHz), 4 GB memory. The performance of the proposed system is evaluated at the pixel level by calculating the true positive rate (TPR) and FPR defined as follows: (7) (8) 5.3 Robustness test To check the robustness and susceptibility of our method, experiments are performed using the above mentioned datasets for the following cases: Copy regions and move without any geometric transformation. Copy regions and move after rotation. Copy regions and move after scaling. Copy regions and move after scaling and rotation. 5.4 Experimental results and analysis for MICC-F220 dataset The experimental results of the proposed system using MICC-F220 dataset images are shown in Figs. 4-6. These results show that the performance of the proposed system is quite satisfactory in the presence of different geometric transformations. We analysed the robustness of the system against rotation, scaling as well as the combination of both and found good detection accuracy. In addition to that, we check the system performance quantitatively by calculating TPR, FPR and computation time in seconds for MICC-F220 images. After experimenting on all 220 images of the MICC-F220 dataset we found that the overall values of TPR, FPR and computational time per image are 99.15%, 3.16% and 3.62 s. respectively. Fig. 4Open in figure viewerPowerPoint MICC-F220 forged images are shown in the first row; the corresponding detection results are given in the second row (a) Without attack, (b) Rotation attack, (c) Scale attack, (d) Rotation + Scale attack Fig. 5Open in figure viewerPowerPoint MICC-F220 forged images are shown in the first row; the corresponding detection results are given in the second row (a) Without attack, (b) Rotation attack, (c) Scale attack, (d) Rotation + Scale attack Fig. 6Open in figure viewerPowerPoint MICC-F220 forged images are shown in the first row; the corresponding detection results are given in the second row (a) Without attack, (b) Rotation attack, (c) Scale attack, (d) Rotation + Scale attack 5.5 Experimental results and analysis for MICC-F2000 dataset The forgery detection results of the proposed system are slightly lower in the MICC-F2000 dataset because the images of this dataset are of high resolution and more complex than the ones in the MICC-F220 dataset. Figs. 7 and 8 show the qualitative results of the proposed system on the images of the MICC-F2000 dataset. The first row of these images shows the input tampered images and the second row shows the corresponding detection results. We analysed the experimental results on different variations of rotation and scaling geometric transformations. In Fig. 7a, no geometric transformation has been applied in the copied region. The copied region of Fig. 7b has gone through 40° rotation. The copied region of Fig. 7c has gone through scaling transformation with scaling factors and . The copied region of Fig. 7d has gone through both the scaling and rotation transformation. Similarly, the copied region of Fig. 8b has gone through 90° rotation. The copied region of Fig. 8c has gone through scaling transformation with scaling factors and . The copied region of Fig. 8d has gone through both scaling and rotation transformation. After performing the experiment on the MICC-F2000 dataset images, the average TPR, FPR and computation time (for one image) are 98.5%, 6.8% and 18.6 s recorded, respectively. Fig. 7Open in figure viewerPowerPoint MICC-F2000 dataset forged images with different attacks are shown in the first row and the corresponding forgery detection results are given in the second row (a) Without attack, (b) Rotation attack, (c) Scale attack, (d) Rotation + Scale attack Fig. 8Open in figure viewerPowerPoint MICC-F2000 dataset forged images with different attacks are shown in the first row and the corresponding forgery detection results are given in the second row (a) Without attack, (b) Rotation attack, (c) Scale attack, (d) Rotation + Scale attack 5.6 Experimental results and analysis for the MICC-F8multi dataset In this section, we analyse the performance of the proposed system in the situation where multiple forgeries are present in tampered images. For this, the experiment is performed on the MICC-F8multi dataset images. In these images, one or more image regions are copied and pasted in more than one distinct location of the same image. To handle the multiple forgeries the generalised 2NN, SIFT keypoints matching procedure and then DBSCAN clustering on the matched keypoints spatial position are utilised in order to separate the different copied regions. Figs. 9 and 10 present the qualitative results of the proposed system using MICC-F8multi dataset images. Different input images are shown in the first row and the corresponding forgery detection results are given in the second row. For the MICC-F8multi dataset images, the average TPR, FPR and computation time (for one image) are 98.8%, 6.9% and 15.6 s, respectively. From these experimental results, it is clear that the proposed system is able to detect accurately multiple forgeries present in images with very few false points detection. Fig. 9Open in figure viewerPowerPoint MICC-F8multi forged images are shown in the first row; the corresponding detection results are given in the second row Fig. 10Open in figure viewerPowerPoint MICC-F8multi forged images are shown in the first row; the corresponding detection results are given in the second row 5.7 Performance comparison of the proposed system with the existing methods The performance comparison of the proposed method with the existing methods in terms of TPR, FPR and average computation time per image is given in Table 2. These methods are tested on the MICC-F220 database images. It can be observed that the performance of the proposed method is superior to all the methods given in Table 2 in all aspects except for the methods proposed in [9,2]. The computational time method proposed in [9] is better than the proposed method but the TPR of this method is very low. Amerini et al. [2] achieved 100(%) TPR using the MICC-F220 dataset, on the other hand, the proposed system achieved 99.15(%) TPR using the MICC-F220 dataset. However, the results of the proposed system are better than that obtained in [2], in terms of FPR and computational time. Therefore, it seems that the performance of the proposed method is quite satisfactory while considering all these aspects. Table 2. Performance comparison of the proposed method with the existing methods in terms of TPR, FPR values (%) and computational time for the MICC-F220 dataset Method FPR, % TPR, % Time, s Fridrich et al. [3] 84 89 294.69 Popescu et al. [24] 86 87 70.97 Bo et al. [9] 3.64 73.84 2.85 Li et al. [25] 8.86 91.55 14.45 Yang et al. [12] 9.02 95.88 10.20 Zhong et al. [17] 14.82 93.75 22.40 Amerini et al. [2] 8 100 4.94 Yang et al. [14] 10.42 95.45 12.40 proposed method 3.16 99.15 3.62 The performance comparison of the proposed system with the existing methods in terms of TPR, FPR and average computation time per image for MICC-F2000 dataset is given in Table 3. From these comparative results, it is clear that the performance of the proposed system is quite satisfactory over the large MICC-F2000 dataset. The computational time and FPR rate of the proposed method are less than the other existing methods. The proposed method also achieved the highest average TPR rate of 98.5%. However, the average TPR over MICC-F2000 is less than the average TPR over MICC-F220. Since MICC-F2000 is the large dataset and contains 2000 large-sized images, therefore, the computational time of the proposed method over this dataset is also higher than the MICC-F220 dataset where the image size is lower than the MICC-F2000 dataset. Table 3. Performance comparison of the proposed method with the existing methods in terms of TPR, FPR values (%) and computational time for the MICC-F2000 dataset Method FPR, % TPR, % Time, s Li et al. [25] 11.8 91.55 24.45 Yang et al. [12] 12.02 92.78 25.20 Zhong et al. [17] 14.82 93.75 22.40 Amerini et al. [2] 11.61 93.42 20.94 Amerini et al. [6] 9.15 94.86 19.20 proposed method 6.8 98.5 18.60 6 Conclusions and future scope In this study, an improved keypoints based CMFD system has been proposed using a density-based clustering approach. Results show that the proposed system performs well in the presence of different geometric attacks like rotation, scaling and composition of these attacks, where the majority of block-based and key-point-based algorithms did not perform well. The robustness of the proposed system is also tested on MICC-F2000 and MICC-F8multi datasets where high resolutions and more challenging tampered images are present. For MICC-F8multi dataset images, the average TPR, FPR and computation time (for one image) are 98.8%, 6.9% and 15.6 s, respectively. From these experimental results, it is clear that the proposed system is able to detect accurately multiple forgeries present in images with very few false points detection. In addition to that, the performance comparison of the proposed method with the eight existing methods shows that this proposed method outperformed the existing methods. Experimental results showed that the proposed method is able to detect small copied regions with the minimum false matches. However, some improvement is needed in the proposed system to improve the forgery detection accuracy in highly similar regions and flat areas since there is no key-point present in flat and similar areas. To overcome the limitations some other feature descriptors can be used with SIFT which are capable of performing flat regions for forgery detection. It is also observed that very few papers have used soft computing techniques in CMFD. The decision-making phase is generally based on determining the values of decision parameters by experience or a result of experiments on a number of forged images. Therefore, the forgery decision can be enhanced or improved by using soft computing techniques in CMFD. 7 References 1Soni B., Das P.K., and Thounaojam D.M.: ‘CMFD: a detailed review of block based and key feature based techniques in image copy-move forgery detection’, IET Image Process., 2018, 12, (2), pp. 167– 178 2Amerini I., Ballan L., and Caldelli R. et al: ‘A SIFT-based forensic method for copy-move attack detection and transformation recovery’, IEEE Trans. Inf. Forensics Sec., 2011, 6, (3), pp. 1099– 1110 3Fridrich J., Soukal D., and Lukas J.: ‘ Detection of copy-move forgery in digital images’. Proc. Digital Forensic Research Workshop, 2003 4Huang H., Guo W., and Zhang Y.: ‘ Detection of copy-move forgery in digital images using SIFT algorithm’. IEEE Pacific-Asia Workshop on Computational Intelligence and Industrial Application, 2008, vol. 2, pp. 272– 276 5Ardizzone E., Bruno A., and Mazzola G.: ‘ Detecting multiple copies in tampered images’. IEEE Int. Conf. on Image Processing, 2010, pp. 2117– 2120 6Amerini I., Ballan L., and Caldelli R. et al: ‘Copy-move forgery detection and localization by means of robust clustering with J-linkage’, Signal Process. Image Commun., 2013, 28, (6), pp. 659– 669 7Pan X., and Lyu S.: ‘Region duplication detection using image feature matching’, IEEE Trans. Inf. Forensics Sec., 2010, 5, (4), pp. 857– 867 8Su B., and Kaizhen Z.: ‘ Detection of copy forgery in digital images based on LPP-SIFT’. Int. Conf. on Industrial Control and Electronics Engineering, 2012, pp. 1773– 1776 9Bo X., Junwen W., and Guangjie L. et al: ‘ Image copy-move forgery detection based on SURF’. Int. Conf. on Multimedia Information Networking and Security, 2010, pp. 889– 892 10Lin S.D., and Wu T.: ‘ An integrated technique for splicing and copy-move forgery image detection’. Int. Congress on Image and Signal Processing, 2011, vol. 2, pp. 1086– 1096 11Mishra P., Mishra N., and Sharma S. et al: ‘Region duplication forgery detection technique based on SURF and HAC’, Int. J. Adv. Comput. Technol., 2013, 6, (17), pp. 1– 8 12Bin Y., Xingming S., and Honglei G. et al: ‘A copy-move forgery detection method based on CMFD-SIFT’, Multimedia Tools Appl., 2017, 77, (1), pp. 837– 855 13Zandi M., Mahmoudi-Aznave A., and Talebpour A.: ‘Iterative copy-move forgery detection based on a new interest point detector’, IEEE Trans. Inf. Forensics Sec., 2011, 11, (11), pp. 2499– 2512 14Yang F., Li J., and Lu W. et al: ‘Copy-move forgery detection based on hybrid features’, Eng. Appl. Artif. Intell., 2017, 59, pp. 73– 83 15Jen-Chun L., Chien-Ping C., and Wei-Kuei C.: ‘Detection of copy-move image forgery using histogram of orientated gradients’, Inf. Sci., 2015, 321, (13), pp. 250– 262 16Xiuli B., Chi-Man P., and Xiao-Chen Y.: ‘Multi-level dense descriptor and hierarchical feature matching for copy-move forgery detection’, Inf. Sci., 2016, 345, pp. 226– 242 17Zhong J., Gan Y., and Young J. et al: ‘A new block-based method for copy-move forgery detection under image geometric transformation’, Multimedia Tools Appl., 2017, 76, (13), pp. 14887– 14903 18Jiangbin Z., Yanan L., and Jinchang R. et al: ‘Fusion of block and keypoints based approaches for effective copy-move image forgery detection’, Multidimens. Syst. Signal Process., 2016, 27, (4), pp. 989– 1005 19Soni B., Das P.K., and Thounaojam D.M.: ‘ Blur invariant block based copy-move forgery detection technique using FWHT features’. Int. Conf. on Watermarking and Image Processing, 2017, pp. 22– 26 20Soni B., Das P.K., and Thounaojam D.M.: ‘ Copy-move tampering detection based on local binary pattern histogram Fourier feature’. Seventh Int. Conf. on Computer and Communication Technology, 2017, pp. 78– 83 21Warif N.B.A., Wahab A.W.A., and Idris M.Y.I. et al: ‘Copy-move forgery detection: survey, challenges and future directions’, J. Netw. Comput. Appl., 2016, 75, pp. 259– 278 22Lowe D.G.: ‘Distinctive image features from scale-invariant keypoints’, Int. J. Comput. Vis., 2004, 60, (2), pp. 91– 110 23Ester M., Kriegel H.P., and Sander J. et al: ‘ A density-based algorithm for discovering clusters a density-based algorithm for discovering clusters in large spatial databases with noise’. Proc. Second Int. Conf. on Knowledge Discovery and Data Mining, 1996, pp. 226– 231 24Popescu A.C., and Farid H.: ‘Exposing digital forgeries by detecting traces of resampling’, IEEE Trans. Signal Process., 2005, 53, (2), pp. 758– 767 25Li J., Li X., and Yang B. et al: ‘Segmentation-based image copy-move forgery detection scheme’, IEEE Trans. Inf. Forensics Sec., 2015, 10, (3), pp. 507– 518 Citing Literature Volume12, Issue11November 2018Pages 2092-2099 FiguresReferencesRelatedInformation
Referência(s)