Research On Fast Algorithm For 3D Video Coding Based On 3D-HEVC

3D video system can provide immersive visual experience for viewers.Recently,the development of display technology leads to a growing popularity of 3D video applications,such as 3D movie,free-view TV,3D home entertainment and virtual reality system.However,compared with the traditional 2D videos,3D videos often contain multiple texture and depth components belonging to different viewpoints,along with some auxiliary information.The huge amounts of data impose heavy burden on the strorage space and transmission bandwidth.To efficiently compress these 3D video data,joint.collaborative team on 3D video coding established by international standardization organizations have launched the 3D extension of high efficiency video coding standard(3D-HEVC),which achieves higher compressionratio over the previous multi-view video coding strandard.3D-HEVC inherits the quad-tree coding structure from HEVC,and adopts some new tools,which significantly improve the coding efficiency,but also result in high computational complexity.Therefore,it is a critical issue to reduce the computational complexity of 3D-HEVC and facilitate its practical application.Based on this background,the research of fast algorithm for 3D-HEVC is carried out.To reduce the complexity of coding unit tree partitioning process,fast coding unit size selection algorithms for base and non-base viewpoints based on spatial-temporal and inter-view correlations are proposed.Base view is coded firstly and does not have reference views for disparity vector prediction.Therefore,hierarchical quad-tree correlations of the current base view can be utilized to skip trying the too large or too small sizes,which could save a lot of time cost for size selection.After coding the base view,warp it onto each non-base view through 3D space transformation,during which hole-marked maps are generated.Then for depth maps in non-base views,coding tree unit partitioning can be early terminated considering the disoccluded information from hole-maked maps;for texture videos,combining the disoccluded information and the inter-view correlations,the coding tree unit splitting process can also be accelerated.In terms of the depth intra coding,a fast depth intra prediction scheme using gray-level co-occurrence matrix(GLCM)is proposed.Before intra coding,GLCMs are generated for each depth coding unit.Firstly,by calculating the correlation feature of GLCMs,the dominate reference orientation for intra prediction can be obtained.Only a few angular modes corresponding to the dominate reference orientation are inserted into the rough mode decision candidate list.Then according to the angular second moment feature and the depth modeling modes usage of adjacent blocks,homogeneous blocks can be detected.And depth modeling modes are omitted to check for homogeneous coding units.Finally,the rate distortion cost calculations are performed for all the candidate intra modes in the list and the ultimate prediction mode can be determined by comparison.The proposed algorithm can reduce the depth intra coding complexity while maintaining comparable coding efficiency.For low-frame-rate 3D-HEVC encoded depth maps,a frame rate up conversion algorithm using graph cuts based motion estimation is introduced.Depth maps in odd-indexed access units are skipped without encoding,and the skipped depth maps can be reconstructed at the decoder using bidirectional motion compensated frame interpolation method.Frame interpolation is conducted within a coding unit tree.The specific block size and search range are determined by motion information of the corresponding texture video.In order to ensure the motion smoothness,the motion estimation process is formulated as global energy minimization function,considering the view synthesis distortion.The function is solved by graph cuts optimization algorithm and the final motion vector can be got.Compared with the virtual views synthesized by normally compressed depth maps,the quality losses of the virtual views rendered by interpolated depth maps are tiny.Additionally,the proposed strategy can save lots of bitrates and encoding time.