Research On Key Techniques Of Forgery Detection For Digital Images And Videos

The extensive application of image/video editing softwares and the corresponding learning resources make the digital images and videos a potential malicious tool, which destroys peoples’ belief in the events described by the visual medias. Such situation motivates the development of the theories and techniques of digital image and video forgery detection, whose purpose is to authenticate the content of given images or videos.In the imaging process of digital image and video, the light reflected by the perceptual objects in the scene undergoes a series operations such as refraction, optical filtering,photovoltaic conversion, demosaicking and post-processing, etc. Each step of the imaging process can be considered as a transform against the light, and an ordered combination of the previous transforms is defined as a chain of transform. In this dissertation, a digital image or video is modeled as a structured combination of some perceptual objects which have undergone certain chain of transforms. On this basis, this dissertation formalizes the image and video forgery operations and attributes the traces caused by forgery operations to two aspects: the existence of excessively similar perceptual objects and inconsistency in chain of transforms. Further, this dissertation models the forgery detection for digital images and videos as a “description and discovery” process: given an input image or video, the detector should firstly find some feature to describe the perceptual objects or certain link of the chain of transforms which the perceptual objects have undergone,to find excessively similar objects or inconsistency in the transform chain, by means of matching or verification. Focusing on the key techniques such as feature construction,matching and verification methods, this dissertation solves several problems in the field of digital image and video forgery detection. The contribution of this dissertation lies in several aspects as follows:Firstly, a feature matching method based on clustering ordered sequence is proposed.Feature matching is one of the key techniques for detecting excessively similar objects. In feature point-based image region duplication detection, existing feature matching methods will miss considerable amount of actually matching feature pairs when several highly similar features simultaneously exist in the feature space. To solve this problem, we propose an algorithm for collecting qualified matching features. The proposed algorithm is based on clustering the elements of an ordered sequence into clusters, and the relevant parameters are adaptively determined by Bayesian Classifier. The completeness of collected qualified matching features can be significantly improved by the proposed algorithm.Secondly, to improve the detection performance of the feature point-based method in smooth regions, we propose a new region duplication detection method based on hierarchical feature detection and feature fusion. Maintaining a moderate amount of total number of feature points, the hierarchical feature detection guarantees sufficient feature point coverage for both the regular and smooth regions. For the feature points located within the smooth regions, by combining a local gradient-based feature and a color feature, we obtain fused local feature which is more discriminative.Thirdly, we propose necessary boundary conditions for DCT coefficient analysisbased compression history inconsistency detection, and devise the corresponding parameter estimation scheme. Inconsistency in compression history between the perceptual objects means that some perceptual objects undergoes different transform in the compression process, compared with the other perceptual objects in the same image. This situation indicates that some region of the image has been tampered. To detect inconsistencies in JPEG compression history, the DCT coefficient analysis-based methods usually use the distributions of the DCT coefficients of the tampered and un-tampered regions as the feature to measure the compression history of each DCT block, and the estimation of the distribution of the DCT coefficients of the tampered and un-tampered components depends on the parameter estimation of a mixture model of the two components. Existing methods estimate the parameters in a blind manner, which usually results in inaccurate estimations. Considering the constraints to which the DCT coefficients should conform,we supplement the likelihood function corresponding to the mixture model with essential boundary conditions. Exploiting the boundary conditions and the smoothness property of the likelihood function, we devise a new parameter estimation method based on coarsegrained search and gradient ascent, which leads to more accurate forgery detection and localization.Fourthly, to detect frame duplication in degraded videos, we proposed a coarse-tofine approach based on the locality sensitive hashing and image registration. Traditional method unexceptionally detect abnormal identical frame pairs in a “feature extraction —thresholding” manner, however, in degraded videos, the local structures of video frames will be changed due to various degradation factors. The accumulation of variation in the local structures of the video frames usually results in substantial change in the features corresponding to the video frames. As a result, a fixed threshold calibrated under certain condition can not stably detect duplicated frames in degraded videos. We pre-cluster visually similar frame sequences by locality sensitive hashing, and validate the potential duplication candidates by frame registration. For better robustness, we encode into the registration energy function the stability information of different regions, and we solve the global optimal matching problem by probabilistic inference.Fifthly, we propose a fast frame duplication detection method for high bit-rate videos.In high bit-rate videos, when there is no significant fluctuation between duplicated frames,we propose the 3D skeleton feature for video frames and the corresponding matching method for fast frame duplication detection. The skeleton feature consists of both topological and geometrical information, which is by nature suitable for a hierarchical comparison. The topological information of the skeleton feature enables us to perform a effective coarse-grained comparison, while the geometrical information is exploited to further identify identical frames.Sixly, we propose a frame deletion/insertion detection method based on detecting abrupt changes in video streams. Many video stream analysis-based methods for frame deletion/insertion detection attempt to detect abnormal periodical artifacts in video stream, however, due to a number of factors, the periodical artifacts can not always be reliably detected. Additionally, existing methods can not accurately locate the position where the tampering operation takes place. In this paper, we propose to use the simultaneous increase of the mean values of prediction errors and number of intra macro blocks of the P frames as evidence of frame insertion/deletion. We devise two features to measure the magnitude of the variation of prediction error and number of intra macro blocks, respectively. Based on the devised features, we construct a fused index to capture the abrupt changes in video streams. The proposed method can effectively detect and locate frame insertion/deletion operations, without any assumptions regarding the encoding process.