A Novel Approach for Copy-move Forgery Detection using Bilateral Filtering

BALKAN JOURNAL OF ELECTRICAL & COMPUTER ENGINEERING, 114 Vol. 8, No. 2, April 2020 A Novel Approach for Copy-move Forgery Detection using Bilateral Filtering N.H. KAPLAN, I.KARABEY AKSAKALLI, U. KILIC, I. ERER  Abstract— Digital image processing methods have a wide area of usage and their complexity is increasing, as well as the tampering methods. A widely used tampering method is copymove forgery. In this study, a hybrid method combining the Discrete Cosine Transform (DCT) and Bilateral filtering is developed. In this method, first overlapping blocks are obtained from the input image. Then, bilateral filtering and DCT of these blocks are multiplied to obtain the refined block features. The block features are scanned by a zig-zag process followed by a lexicographic sorting. Finally, a similarity detection by a predetermined threshold parameter is applied to detect the forgery. Both visual and quantitative results demonstrated that the proposed method can determine the copy-move forgery regions. Index Terms— copy-move forgery, bilateral filtering, zigzag scanning, DCT (Discrete Cosine Transform) I. INTRODUCTION N OWADAYS, digital images are used in important areas such as medical, law and public. Digital images can be manipulated and regulated easily by malicious people using various image regulation software tools. With the emergence of this software, the reliability of the images and their authentication have become an important problem. Therefore, image fraud detection has become an important research focus. Image fraud detection methods are generally divided into two categories as active and passive approaches. Active NUR HUSEYIN KAPLAN, is with Department of Electrical Engineering University of Erzurum Technical University, Erzurum, Turkey,(e-mail: ). https://orcid.org/0000-0002-4740-3259 ISIL KARABEY AKSAKALLI, is with Department of Computer Engineering University of Erzurum Technical University, Erzurum, Turkey, (e-mail: ). https://orcid.org/0000-0002-4156-9098 UGUR KILIC, is with Computer Engineering University of Erzurum Technical University, Erzurum, Turkey, (e-mail: ). https://orcid.org/0000-0003-4092-3785 ISIN ERER, is with Department of Electronical and Communication Engineering University of Istanbul Technical University, Istanbul, Turkey, (email: ). https://orcid.org/0000-0002-2225-6379 Manuscript received November 27, 2019; accepted Feb 6, 2020. DOI: 10.17694/bajece.651435 Copyright © BAJECE approaches are based on additional information embedded into digital images such as digital watermarks or digital signatures. By using these additional information, the originality of the image can be detected. Unfortunately, active approaches require additional information to be embedded in the image by authorized personnel in the process of capturing the false image or at a later stage. If there is no information about the original image, applying an active approach is not useful [1]. On the other hand, passive approaches are used to determine the manipulated image without any additional information. Passive approaches are divided into two groups named tampering detection and source device identification. This approach detects copied image by extracting real features in the image. Tampering detection is also divided into dependent and independent classes. The dependent class of copy-move forgery is the commonly used method in fraud image. The image content is manipulated by copying an object that exists in the image and paste this object to another location within the same image. The transition between copied object and original image is masked using a variety of retouching tools. Since the features such as noise, color and contrast in the source and target regions have a statistical match, the detection of copied zones is a challenge. Fridrich et al. [1], pioneers of copy-move forgery detection algorithm (CMFD), has handled the various requirements of the detection algorithm. The first requirement is that the detection algorithm should allow the approximate matching of the small image segments. Secondly, while the detection algorithm determines the mismatched fields (false positive), it must have an acceptable execution time. Furthermore, the authors mentioned that a fake segment will likely have a dependent component rather than very small patches or individual pixels. In this study, a novel method consisting of a combination of Discrete Cosine Transform (DCT) and zigzag scanning is proposed by applying bilateral filtering. DCT is one of the most used watermarking algorithms among many data hiding methods to protect digital multimedia files. It states a limited sequence of data points in terms of a sum of cosine functions in different frequencies. Bilateral filtering has been used in many image processing algorithms [2-4]. Bilateral filters take into consideration both spatial and spectral properties of the image. By this way, the edge information is kept during the filtering process. The method is compared with traditional and state of art methods and the proposed method gives 99.7% accuracy rate in a standard dataset called “CoMoFoD” used for benchmarking the detection of tampering or copied images. ISSN: 2147-284X http://dergipark.gov.tr/bajece BALKAN JOURNAL OF ELECTRICAL & COMPUTER ENGINEERING, II. RELATED WORK It is known that there is a correlation between the original image and the pasted object in copy-move forgery [1]. This correlation is used to detect forgery successfully. The methods for approximate matching of copy-moved and real segments using retouching tools or other image processing tools are not fully sufficient to detect forgery, so various approaches for detecting the forgeries are increasing day by day. In most of the methods proposed for CMFD, the basic procedure is divided into pre-processing, feature extraction, matching, filtering and post-processing stages, respectively [58]. In the pre-processing stage, image data is improved to enhance image features or to reduce undesirable distortions within the image. One of the most commonly used methods at this stage is the conversion of RGB color channels of the input image into a single grayscale image [1], [5-9]. Besides, if the image is stored in a compressed format, the files are decompressed in the preprocessing phase [3]. The feature extraction phase is then applied to the selection of related information representing the properties of the image. This phase is carried out in two ways: dividing into blocks and keypoint detection. In block-based approach, the image is divided into blocks in a square or round shape. These blocks can be divided by overlapping or non-overlapping division [8]. Then, properties are extracted from these blocks by using various features (frequency transform, texture and intensity, moments invariant, log-polar transform, dimension reduction etc.) and similarity comparison is performed between the blocks in the image. The third stage, matching, is a (...truncated)