Projection Kernel Method In Image Tamper Detection

With the wide spread of image processing software and the rapid development of multimedia technology, digital images can be easily manipulated and altered. Abusive use of digitally forged images has become a serious problem in various fields such as news media, academic research, law, militaries, and etc. Image blind forensics emerges under this background and aims to determine the authenticity, integrity and originality of the digitally image without relying on any auxiliary information. This thesis makes an in-depth research analysis on image forgery detection based on image noise characteristics, artificial blurring trace and composite forgery. The main contents and innovations are as follows:As to the study of image noise detection, an analysis on the limitation of current model for image noise detection is presented. Since the traditional methods require training image database or prior information, it restricts their practicable application. In order to solve this problem, we propose a new method of image noise detection which can be implemented using the general vector space projection or a new tensor space projection. First, we use the noise subspace, which is constructed by principal component analysis(PCA), two-dimensional PCA(2DPCA) or kernel PCA(KPCA), to detect and locate the forged regions. Second, tensorial two-dimensional PCA(T2DPCA), a novel approach considering the third-order tensors as linear operators on the space of oriented matrices, which has been proposed recently and showed better performance than traditional PCA in image analysis and recognition, is used to construct the noise subspace. Feasibility of the addressed methods is illustrated by experimental results on image forgery detection.In order to solve the problem of the space complexity and effectively extract nonlinear structure from the forged images, T2 DPCA is further extended to an incremental version and a kernelized version with better performance.In the first step of the above extension, an effective incremental method for performing T2 DPCA in the tensor spaces is proposed by extending Oja’s rule to tensorial version and a related sparse representation technique is thus obtained by a tensor-based distance criterion. First, we directly extend Oja’s rule by replacing the general vector product with the defined tensor product, called T-product, in the third-order tensor spaces. Then, the extension to multiple principal tensors is straightforward by using modified inputs. Second, we consider a tensor-based distance criterion to provide sparse representation. We propose incremental T2DPCA(IT2DPCA) by combining the above two steps.While in the second step, by assuming that an eigenvector of the covariance tensor of the 2D training images is the tensor linear combination, called T-linear combination, of the training images, the T2 DPCA is improved to a new version. And the improved method is further extended to a nonlinear version called kernel T2DPCA(KT2DPCA), by using the general kernel trick in machine learning field, but with a new inner product called inside product which is defined with the T-product in the tensor spaces, and togther with the general inner product defended in vector spaces.And in the last step, in a similar way, we give a nonlinear derivation of IT2 DPCA, called incremental KT2DPCA(IKT2DPCA). Experimental results showed satisfactory performance of the proposed methods.As to the study of blurring trace detection, firstly the existing methods are studied and analyzed. Then, under the perspective of the traditional idea, we propose a new method of blurring trace detection based on the linear correlation of image pixels, which can be extracted by using least square(LS) method or singular value decomposition(SVD). Experimental results show that the proposed methods could reveal blurred regions that indicate possible tampering.In addition, since most existing image forgery detection methods consider only one single feature of blurring operation, we propose to adopt feature fusion involving multifeatures for blurring operation in image tampering to improve the accuracy of forgery detection. First, three feature vectors that address the singular values of the gray image matrix, correlation coefficients for double blurring operation, and image quality metrics(IQM) are extracted and fused using PCA or KPCA, and then a support vector machine(SVM) classifier is trained using the fused feature extracted from training images or image patches containing artificial blurring operations. Finally, the same procedure of feature extraction and feature fusion are carried out on the suspected image or suspected image patch which is then classified, using the trained SVM, into forged or non-forged classes. Experimental results show the feasibility of the proposed method for image tampering feature fusion and forgery detection.As to the study of the composite forgery detection, the existing information fusion methods of image forensics are reviewed and analyzed. In most of the currently presented methods, the related image forgery detection is usually carried out by using feature fusion or decision fusion. We propose a hierarchical fusion framework consisted of feature fusion and decision fusion to improve performance accuracy in image forgery detection. In the framework, multiple inherent features of a suspected image are firstly extracted and fused by using kernel discriminant analysis(KDA), and then classified into forged, non-forged or undetermined classes. For a undetermined image, tampering features are extracted and fused by using evidence theory to fulfill detection task. Experimental results show the feasibility of the proposed method for image forgery detection.