Copy-Move Forgery Detection Algorithm For Digital Images

Digital images are becoming more and more important as the main carrier of information.However,with the popularity of image acquisition devices and the rapid development of image editing software,in recent years,digital image fraud events have emerged in an endless stream,which not only reduces the credibility of digital images,but also has a great negative impact on society and individuals.Image copy-move forgery is one of the most common types of image tampering.It has the characteristics of easy operation and being effective,and is often used to change the semantic information of digital images.This thesis aims to protect the image content by studying the image copy-move forgery detection method.In view of the excellent learning and analysis ability of deep learning,two copy-move forgery detection methods based on deep learning are proposed.The traces left by the image processing operations are used to distinguish the tampered region from the original region.A series of experimental results verify the rationality and detection accuracy of the method.The main work of the thesis includes:(1)This thesis proposed a deep learning based method to detect digital image copy-move forgery and localize the tampered regions.This method uses the pre-processing operation to suppress the influence of image content,and at the same time,the image-blocking operation separates the real region from the tampered region,which successfully changed the learning goal of the convolutional neural network to identify local regions with inconsistent CFA characteristics in the image.Thereafter,a convolutional neural network with only one convolutional layer is designed and trained to learn the fringe(or grid)-like pixel value distribution characteristics on the real area of prediction error images.Finally,the detection and localization accuracy of the method is further improved by eliminating the false detection image blocks in the background area and repairing the missing detection blocks in the detected area.The experimental results show that the proposed method achieves a detection accuracy of 97%at the block level and a localization accuracy of 80%at the pixel level.(2)A method based on deep learning is proposed for distinguishing different image tampering operations.The problem to be solved is to distinguish the rotation tampering area,the scaling tampering area and the real area in the image,which is a three-class problem.The method follows the detection framework and convolutional neural network structure of the first work.Based on this,two aspects are improved:First,a plurality of filtering methods are used to obtain a set of residual images.This method provides convolutional neural networks rich and diverse data to promote their classification performance.The second is to use the weighted voting strategy to synthesize the classification results of multiple networks,thereby accumulating the advantages of each model,breaking the limitations of single model,and thus improving the performance of the method.Finally,the integrated classification results are corrected to obtain more accurate test results.The experimental results show that the proposed method can effectively locate the rotation and scaling tampering regions,and the macro average and micro-average reach 91.9%and 91.8%,respectively.