Research Of Image Passive Forensics Based On Statistical Modeling

With the rapid growth of digital devices and image editing software, the digital image has become easier than ever to modify. If the important image is tampered maliciously, it will have a bad effect on society. Compared with image active forensic technology like digital watermarking, image passive forensics do not need any prior information, and has wide application value. This paper mainly researched on the methods of image passive forensics based on statistical modeling. With the use of machine learning tools, it proposed a new image forgery detection method based on the statistical model that recently proposed in the steganalysis field. Since the image splicing tampering will change the texture of image, it proposed an improved statistical model based on texture analysis method to extract the feature vectors and successfully applied for splicing forgery detection. This paper aims at finishing the work below.(1) Studying the related methods of image statistical modeling. First, it analyzes the relationship between steganalysis and image forgery detection. It also researches the relevant statistical models that has been successfully used in the steganalysis field. Then, it studies the image texture analysis methods that used very often. It emphatically introduces the local binary pattern(LBP) technology and the co-occurrence matrix technology, both of which can give the compact and roust statistical description. Then it introduces the linear and nonlinear highpass filters that used to calculate different residual images. Note that the extract operation of statistical features is built on the basis of the residual images. Finally, it gives the introduction of two kinds of machine learning tool, the support vector machine(SVM) and ensemble classifier.(2) Researching the feature extraction model of residual-based local descriptors and its application in image forgery detection. First, it has a deep study of the feature extraction method of residual-based local descriptors that recently proposed in the steganalysis field. Then, the residual-based local descriptors will be applied to image splicing forgery detection and image recapture forgery detection with the use of machine learning tools. During the experiments of splicing forgery detection, according to the obtained estimate value of AUC of each residual-based local descriptors, it tries to merge the descriptors that have higher AUC values. During the experiments of recapture forgery detection, greedy method is utilized as the combination strategy, namely each time using exhaustive search to select a feature vector that can obtain best detection accuracy after it being merged. Since the type of tampering is diversity and residual-based local descriptors may fail to detect the forgeries that are too small, at the advanced stage of the proposed solution, it also conducts a image copy-move forgery detection using block matching method based on the Patchmatch algorithm to the images that were judged as pristine by previous method.(3) Researching the image splicing forgery detection based on LBP and co-occurrence matrix. Due to the problem that splicing tampering will cause texture changes of image, it proposes a method for splicing forgery detection based on image texture analysis on the basis of the feature extraction model of residual-based local descriptors. This method combines the local binary pattern(LBP) and the co-occurrence matrix two effective texture analysis technologies. Two improved LBP methods are utilized to describe the local texture of the image after residual images were computed. Then it uses two types of co-occurrence matrix method to extract various feature vectors. In order to improve the detection accuracy, these feature vectors are further assembled using some strategies. Support vector machine(SVM) or ensemble classifier is utilized as the classifier in the proposed method.