Multi-view Video Coding Based On Stereo Perception

Three Dimensional (3D) video consisting of two or more views is able to show a realistic andinteractive scene. As the efficient3D data representation, Multi-view Video plus Depth (MVD)system must capture a very large number of views compared to2D video. So, methods toefficiently compress stereoscopic video have been actively studied. Moreover, understanding theperception characters of Human Visual System (HVS) plays the key role in improving theefficiency of video compression. In this thesis, perceptual multi-view video coding is devised.(1) MVD consumers more bitrate than the2D Television. Combined coding results of thevideo and depth data are analyzed to determine the relationship of Quantization Parameters (QP) inthe video and depth data. Then how the separate coding results influence the view synthesis qualityis investigated in the form of the left and the right view, as well as the edge regions and thenon-edge regions in the depth data. Finally the joint coding for MVD is proposed. The proposedalgorithm can achieve a better RD performance compared with the temporal and inter-view skipmodes for multi-view video coding of the Lee’s and the virtual view oriented distortion criterionfor depth maps coding of Yuan’s.(2) The foveated just noticeable distortion model based multiview depth maps coding todecrease the bitrate in depth coding is proposed. The foveated just noticeable distortion maps in theintermediate view is firstly estimated. Then, the allowance-maximum distortion in the depth mapof left view is calculated. As for the depth maps of right view, it is coded by adopting therelationship of the quantization parameters of the left view. The proposed algorithm, which savesthe bitrate by19.44%~34.12%, is more prominent than the virtual view oriented distortion criterionfor depth maps coding of Yuan’s.(3) A fast mode selection of multi-view video coding based on the binocular just-noticeabledifference model is proposed to decrease the complexity. Firstly the disparity between two views isutilized to obtain the BJND in the right view. Then the statistical relationship between the BJNDand the mode selection is established. Finally, the fast algorithm half stops the process ofmacroblock mode selection. The proposed algorithm saves the encoding time by70.06%whilemaintaining almost the same RD performances as the original encoder.