Study On Passive And Blind Digital Image Forensics Using Characteristics Introduced By Image Formation

The popularization of digital camera and the wide spread of sophisticated image processing software, make image forgery creation easier and easier in the digital era. In recent years, people show deep worries about the authenticity of multimedia, because of the frequent disputes caused by the falsification of digital images in various areas such as politics, economy and news. As a result, digital image forensics, which concerns the integrality and authenticity of images, receives considerable attention these years. Consequently, the passive and blind digital image forensics, which only uses the image for analysis, is becoming a hot topic for the academic research. Aiming at the various problems in this field, a study on passive and blind digital image forensics has been made on the aspects of source camera identification of digital images, splicing localization, and blur localization, based on the features introduced by image formation. The main contribution of this dissertation is as follows:As to the source camera identification, an analysis on the limitation of current model for source camera identification is presented firstly, and then two problems are pointed out. One is the decreasing accuracy of the source camera identification as the number of the camera models increasing. The other is that the existing methods are incapable of the outlier models of the training model. Under the perspective of digital image formation in digital camera, CFA interpolation, the key operation in image formation, is utilized for source camera identification. A method for estimating CFA interpolation coefficients is proposed based on covariance matrix. Furthermore, two algorithms, which are based on multi-class classifier and the combination of multi-class and one-class classifier, for source camera identification are presented respectively using the CFA interpolation coefficients as feature vector.As to the study of splicing localization, the existing methods of splicing detection are reviewed and analyzed. Furthermore, a general model of splicing localization for spliced forgeries is proposed. By estimating the original image counterpart, the coherence and consistency in the neighborhood of the pixels, which may destroyed by splicing, are recovered for localizing the tampered area. Based on this theory, a framework using'soft imaging' is presented for original image counterpart estimation, and three approaches of soft imaging, which are CFA interpolation, JPEG compression and adaptive CFA interpolation, are proposed respectively, to localize the splicing area in pixel-level. As to the study of blur localization, an observation is firstly pointed out that the coherence and consistency exist in color space in the neighborhood of the pixels introduced by the image formation, and also a theoretical analysis of the effects caused by blur tampering is made. On the basis of these work, the abnormal hue in HSI color model is proposed to describe the coherence of the color space in the neighborhood of pixels. Moreover, two methods based on abnormal hue and hue characteristics are presented to localize the blur tampering.