Blind Separation Method For Permuted Alias Image Based On Differential Evolution

Blind source separation, one of the most hot research directions in modern signal processing field, is a procedure to separate the different source signal from many mixed signals under the circumstance that both the source signal of the observed one and the mixed model are unknown. It is widely used in image tampering detection. In image tampering, a special single channel alias image called as permuted alias image is encountered. Because of the essential difference between transfer mode for permuted alias image and hybrid approach for the normal single channel signal, permuted alias image can’t be accurately separated by the theory of blind separation for single channel signal. But most of the existing blind separation methods for permuted alias image focus on the concrete methods of feature extraction and lack of integrity and feasibility. In this paper, the research is carried out to solve the problem of lacking of integrity in the existing blind separation for permuted alias image. The key to complete separate permuted region from the permuting part is to find out the difference between features and appropriate separation limits. Based on the separation boundary obtained by differential evolution algorithm, a complete blind separation method for permuted alias image solution was proposed.The primary contributions of the research are summarized as follows :(1) Through the research of the blind separation for permuted alias image, on this basis, a new modified blind source separation scheme based on dynamic threshold is proposed.(2) On the premise that the permuted image has been interpolated, this thesis presents a new blind separation theory for interpolation permuted alias image. By analyzing the features of the interpolated images, the differences between them and images without interpolated are found. Then the finite difference method is used to present them. After this, set threshold to the sub-blocks and form them to a threshold vector. The optimal threshold is obtained by finite difference optimization algorithm if the function is chose properly. Thresholding the differential image into binary one. Then the permuting image is separated by doing a dot products between the source image and binary image. The experimental results demonstrate that, under the case that all of the prior knowledge is unknown, this algorithm has the ability to accurately separate the permuted region while indicating a stronger robustness.(3) This paper proposed a complete blind separation theory for the permuted alias image with noise. Noisy images can’t be represented by sparse representation while natural images without noise can. According to this fact, KSVD dictionary learning is selected for sparse representation. The permuting region is obtained by the difference between the denoising image and the source image. After dividing the difference image into sub-blocks, the sub-blocks are set threshold. Form the value into a threshold vector. Under this consideration of global image and the part region of image with noise, the target function is set. The optimal threshold which is obtained by differential evolution will separates the permuted image. The experimental results demonstrate that, under the case that permuted region’s size, location, quantity and noise standard deviations of the permuted region are different, this algorithm can accurately locate and separate the permuted region and indicate a stronger robustness and get a better performance for various interpolation methods.