Research On Side Information Generation For Multi-View Distributed Video Coding

Distributed Video Coding (Distributed Video Coding, DVC) is the Slepian-Wolf and Wyner-Ziv (WZ) theory-based video coding technology. Unlike the traditional video coding methods, the paradigms of DVC move the complex prediction techniques from the encoder to the decoder, particularly suitable for terminal devices with limited coding capacity. In order to get better compression performance, only parity bits are transmitted to the decoder after the WZ frame been encoded. While the decoder is lack of the original WZ frame, a corresponding prediction version called side information (SI) need to be created. The accuracy of SI will influence the total DVC system. In recent years, many effective SI generation methods have been existed in single-view field, but in multi-view field, the spatial SI generation method is still lack of efficacy.This thesis focuses on side information generation for Multi-view DVC. The results obtained are as follows.Firstly, a spatial side information generation method for Multi-view DVC based on hybrid search and multi-selection is presented. The spatial SI generation method combined feature matching and epipolar line search to improve the matching accuracy, then obtained the optimum location by multi-selection. The average value of peak signal noise ratio (PSNR) by applying the proposed method is higher than the traditional method disparity compensated view prediction (DCVP) about 1.65 dB in the best case, and 0.64 dB in the worst. The proposed method reached the similar results compared with disparity based view synthesis (DBVS),without any information about the camera intrinsic parameters.Secondly, a fusion mask generation method for side information fusion has been proposed. The proposed method generate the temporal binary mask by comparing the backward and forward frame in temporal with the temporal SI, then generate the spatial binary mask by comparing the left and right frame in spatial with the spatial SI, finally a pixel mask based on spatial-temporal degree of confidence can be created by the motion vector and disparity vector from the SI generation module. After combining with all the masks, the fusion mask is created. The proposed fusion method can improve the rate-distortion of the Multi-view DVC system compare to the traditional temporal based or space-time compensation based fusion methods, can select and save the details in the temporal and spatial SI to enhance the reconstruction quality of the final decoded image.