Research On Correlation Noise Model In DCT Domain Distributed Video Coding

The Distributed Video Coding (DVC) is a new paradigm of video compression,unlike the conventional video coding (MPEG-x, H.26x, etc.), it shifts computationalcomplexity from encoder to the decoder, and applies channel coding techniques.Therefore the DVC system has the characteristic of simple encoding, complexdecoding, and high compression efficiency. Thus, DVC is especially suitable for thewireless video applications where computing power, storage capacity and powerconsumption are limited. The coding efficiency of DVC codec heavily depends on themodeling accuracy of the correlation noise between corresponding DCT bands of theoriginal Wyner-Ziv (WZ) frames and the side information. Therefore this papermainly focuses on the correlation noise model in DCT domain Distributed VideoCoding. The main work and achievements are as follows:1. This paper systematically introduces the basic principles and key technologiesof Distributed Video Coding (DVC), and then implements the classic DCT domainDistributed Video Coding based on Turbo Codes. And then the RD performance ofDVC is compared with the H.264/AVC intra-frame coding.2. Because the quality of the same side information frame is different with itsspatial location, so the channel noise characteristics are not the same in differentregions. In the DCT domain, the channel noise characteristics are also different in thedifferent location area of the same band. According to this feature, this paper presentsa correlation noise model estimation method based on the K-means clustering. In themodel, the DCT residual coefficients in each band are first clustered into three sets.Then the residual coefficients in each set are utilized to calculate the Laplacianparameter. Finally the corresponding Laplacian model is obtained. The experimentalresults show that the algorithm can more accurately describes the distribution of theresidual coefficients between WZ frame and side information, and effectively improveR-D performances of the Distributed Video Coding.3. According to the characteristics of the DCT residual coefficients between corresponding DCT bands of the WZ frame and the side information, this paperpresents a novel Laplacian–Cauchy Mixture Distribution (LCMD) model by modelingsmall DCT coefficients as improved Laplacian distributed while modeling large onesas Cauchy distributed. Two solutions to find parameters of LCMD model are alsoproposed. Real video experiments demonstrate the improvements of LCMD in termsof both the RD performance and the required decoding CPU time.