Researches On Fast Algorithm For Multivew Color And Depth Video Coding

As the telecommunication technology emerges rapidly, the industry and consumer electronics field are more and more interested in multiview video systems, such as three dimensional television (3DTV), free-viewpoint television (FTV) and photorealistic rendering of three-dimensional (3D) scenes. Multiview view plus depth (MVD) is a mainstream to represent 3D scenes. MVD, as one of the most important multiview video data formats, includes multiple viewpoint color videos and depth videos. It fulfills the 3D video system's requirements and supports wide-viewing angle of 3D displays and auto-stereoscopic displays. Moreover, it also allows rendering a continuum of output views with high image quality. In order to store and transmit the huge data of MVD, some technologies, such as view scalability, multiple reference frames, motion estimation and disparity estimation, are utilized in multiview video coding (MVC) system. However, high compression efficiency is achieved at the expense of increased computational complexity. This thesis mainly focuses on fast algorithms for multiview color and depth video coding and depth prepropcessing method. And the contributions of this thesis are:(1) This thesis proposed a fast multi-reference frame selection algorithm based on dynamic threshold after analyzing the statistical features of multi-reference frames and the distribution of best search detection in MVC. Dynamic threshold technique is given to terminate the process of searching multi-reference frames early. The proposed algorithm also reduces the useless candidate search detections which re less likely to contain the best matched block. Experimental results show that the proposed algorithm can achieve 50.25%~72.22% reduction of encoding time in comparison with Joint Multiview Video Model (JMVM), while average bit rate increase and PSNR degradation is within 0.54% and 0.06dB for test multiview sequences.(2) This thesis proposed an edge-attribute fast motion and disparity estimation algorithm. Firstly, the conception of edge-attribute is defined, and the regular pattern of different blocks with different edge-attribute is analyzed. In MVC, the distribution of motion vector (MV) and disparity vector (DV) is regular. The blocks with same edge-attribute have similar MVs and DVs. Then, an MV and DV prediction method for increasing the accuracy of motion and disparity estimation is proposed. By utilizing the proposed MV and DV prediction method, the best matching block can be found exactly in small search range. Therefore, the computation of motion and disparity estimation in inter-view channels can be greatly reduced. The fast algorithm takes advantage of refinement-search method to keep the best quality of coding video. Experimental results show that the proposed algorithm speed up 17.84~31.40 times compared with the full search in JMVM.(3) This thesis proposed a novel depth preprocessing method based on 3D visual perception to revise the inaccuracy of depth map. Firstly, we define perception factor according to 3D visual perception depth sensation of human visual system. Then, the perception factor is utilized to decide which part of depth map should be protected and which part need to be smoothed by low-pass filtering in temporal direction. And then, smooth the depth video with aid of color video. Experimental results show that the proposed method saves the bit-rate by 7.02%~48.34% in depth compression while it keeps the virtual view rendering performance.(4) This thesis proposed a fast algorithm based on the inter-view correlation for multiview depth video coding. All the views can be divided into two categories, that is key view and assistant view. Key view is the view in which there is no disparity estimation or only anchor frame have disparity estimation. Assistant view has motion estimation and disparity estimation. Firstly, this algorithm analyzed the temporal correlation of views and the inter-view correlation of assistant views. Then, the assistant view encoded can predict the inter-view correlation of current assistant view. Experimental results show that the proposed method can save 35.90%~63.10% reduction of encoding time while maintains high virtual rendering performance in comparison with full search algorithm.