Study Of The Key Theories And Algorithms On Image Forensics

Digital images have arisen in all fields of our daily life and their influences have been growing. However, the image processing software brings a serious crisis of confidence to digital images. Digital watermark is the representative technique of active protection methods, which rely on certain pretreatments, while in reality, we need to deal with completely unknown images. Therefore passive protection techniques are proposed to solve the crisis of confidence. As an important branch in passive protection techniques, image forensics under blind environments attracts much attention and becomes a hot spot at present.In this paper, by analyzing and investigating the existing image forensics methods, we proposed the improving schemes some aspects such as theoretical innovation and algorithm improvement. The main achievements are as follow.1. We put forward forensics algorithms based on multiple JPEG compression features. We analyzed the existing forensics methods using multiple JPEG compression features, and proposed a zoomed window detecting algorithm. Existing methods detect the abnormal areas according to the features of a single DCT transform grid, so they cannot provide clear boundaries of the distortion. By virtue of the blocks composed with several grids, our algorithm can easily detect the characteristics of the abnormal areas. With a zoomed window, the algorithm can locate the distortion quickly and effectively. Besides, by comparing the compression of the normal style with that of the malicious style, we concluded that the malicious compression will lead to dislocation or deletion on the DCT coefficient. Thus, we put forward a sliding window detecting algorithm to detect the malicious compression. The algorithm can estimate the compression history of the image, and testify the malicious compression.2. We put forward forensics algorithms based on computer vision. As one of the basic tools in computer vision, the camera calibration tools are frequently used in image forensics. By using camera calibration tools, we can estimate the principal point, which can help verify the image. However, camera calibration tools have some limitations in image forensics since they can only work when there exist images in mulitple views or konwn cameras. Therefore, we put forward a multi-perspective scheme to solve the problem, and proposed a forensics algorithm according to regular graphics. Compared with the original algorithms, our algorithm can deal with completely unknown images, and the features we used are robust to smoothing, noising and JPEG compressing. Moreover, based on the orthometric one-dimensional objects and vanishing points, we proposed forensics algorithms, which possess more general applicability and higher accuracy in image forensics. Besides, on the basis of the external matrix, we put forward a method to detect small displacements.3. We proposed forensics algorithms for low resolution images and texts. By analyzing the formation features of low resolution images, we put forward forensics models based on generalized camera model. We summarized the generalized camera model for low resolution by combining the imaging procedure and the compression procedure. Then we transplanted our algorithms to detect forgeries in low resolution images. For images with texts, by virtue of the imaging features, we located the abnormal texts by reconstructing the texts.4. We reformed some existing methods. We improved forensics methods for texts by combing texts model and abnormal edge features, and put forward a texts forensics method using self-learning system. By introducing KKT constraint, we improved montage detection methods based on the light sources.5. We proposed an authentic assessment model. We classified the alterations according to the impact of the alteration rather than the "alteration=falsify" model. We divided the alterations into three categories:subversive alteration, severe alteration and general alteration. We sorted the existing image forensics methods according to the classification model of the alterations, and according to the classification above, we proposed a hierarchical model for authentic assessment, and we put forward a procedure for the evaluation of news images.