Research On Fast Multiview Video Coding Based On Mode Decision

Multiview video shows a broad application prospect in3DTV, free viewpoint TVand other related fields. Because of its superior telepresence and interactivity to2Dvideo, it has gradually become a hot topic of the multimedia information industry andacademia. But with many sets of adjacent video camera shooting from different angles,the great amount of multiview video data must be effectively compressed to overcomethe technical difficulties of transporting and storing. In order to get high compressionefficiency for multiview video, in July2008, the Joint Video Team (JVT) released themultiview video coding (MVC) standard H.264/MVC as the high profile forH.264/AVC and developed a specific reference model JMVC (Joint Multiview VideoCoding).This thesis gives an in-deep investigation on the fast macroblock (MB) modedecision algorithms for MVC based on JMVC platform, and the major techniques andcontributions are listed as follows:1. The basic principle and structure of H.264/AVC video coding are deeplyanalyzed and the coding framework and coding process of JMVC are studied; thecurrent multiview video coding structures are analyzed and compared where theAS_IPP structure which shows great coding performance is selcected in this research;the mode decision process which is the most time consuming process in multiviewvideo coding is investigated deeply as the foundation of the following multiviewvideo coding research.2. A SVM based macroblock mode decision algorithm for MVC is proposed inthe thesis. Through the analysis and summary of the macroblock mode distribution formultiview video, support vector machine (SVM) which has shown the superiority inthe field of pattern recognition is introduced in MVC in order to overcome the lowrobustness of empirical coefficient in some threshold methods. According to theprinciple of the strong correlation between the information of coded reference viewand the current view, the focused mode decision problem can be converted to a patternclassification problem. The artificial intelligence learning machine is utilized to buildthe classifier model and the low complexity MVC fast mode decision algorithm isproposed to significantly increase the speed of multiview video coding.3. A depth information based mode decision algorithm for MVC is proposed to speed up the MVC process. In the thesis, the relationship between MB mode anddepth information is analyzed and all the MBs in multiview video sequences aredivided according to the corresponding depth value in depth map into remote, closeand midrange areas. Each area is handled respectively. For the midrange area which isthe most complicated one, firstly the most frequently used mode of the correspondingMB and its surrounding MBs in reference view are used as the most likely used mode(MLM) to separate the MBs which may use large partition mode. Then the dynamicdepth flatness combined with the motion information method is used to determine thefinal macroblock mode. The experimental results show that this method cansignificantly reduce the multiview video coding complexity. Meanwhile, it can keepthe same rate distortion performance.