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

In recent years,multimedia communication technology and display technology have developed rapidly,and the application scenarios of 3D video in people's lives has become more and more diverse,such as 3D movies,3D home projectors,3D touch all-in-ones,and VR experience systems.Compared with 2D video,in addition to multi-view texture maps,3D videos also contain multi-view depth maps with more distance and auxiliary information,which has undoubtedly led to an increasing amount of data in 3D videos.In order to solve the problem that 3D video occupies a large transmission bandwidth and storage space,the International Collaborative Team On 3D Video Coding Extension Development(JCT-3V)developed the 3D-HEVC(3D-High Efficiency Video Coding)on the basis of the traditional 2D video compression standard HEVC(High Efficiency Video Coding).3D-HEVC inherits the intra-prediction modes and coding unit quad-tree segmentation structure from HEVC,which greatly improve the compression efficiency of 3D video texture maps.The most obvious feature of 3D video depth map is that it has large areas of smooth areas and sharp edges.The original intra prediction mode of HEVC cannot properly represent the sharp edges of depth map.In order to improve the compression quality of the sharp edges of the depth map,3D-HEVC uses a depth modeling mode for encoding depth map,but at the same time,it increases the complexity of the entire encoding process.The quad-tree segmentation has always been one of the most computationally complex when coding texture maps.Depth map coding has more redundant calculations and recursive calculations of quad-tree segmentation than texture map coding.In order to improve the coding efficiency of the 3D-HEVC depth map in intra prediction mode selection and quad-tree partitioning,this paper has made some optimization:1?A DMMs mode skip algorithm for depth maps based on wavelet transform is proposed.Because the high-frequency coefficients of the wavelet matrix can well reflect the degree of discreteness of image,the algorithm generates the corresponding wavelet matrix before the intra prediction of the depth map prediction unit.Whether the current prediction unit contains sharp edges is determined by the high-frequency coefficient of the wavelet matrix.If the sum of high-frequency coefficient is 0 we can judge that the current predict unit contains sharp edges.If there are no sharp edges,we can directly skip the selection of the depth modeling mode and the corresponding rate-distortion calculation.Thereby the algorithm can reduce the computational complexity of the depth map in the intra-prediction encoding process.Experimental results show that compared with the original HTM13.0,the algorithm reduces the overall coding time by 4.6%under the same video quality.2?We proposed an algorithm based on wavelet transform for early termination depth map quad-tree segmentation.Firstly,determine whether the current prediction unit is absolutely smooth by the high-frequency coefficients of the generated wavelet matrix and whether it is 0.Secondly,determine whether the current coding unit is relative smooth by calculating whether the average value of pixels around the prediction unit is less than the given threshold.For coding units that satisfy absolute smoothness and relative smoothness,they can directly jumped out the recursive operation of quad-tree division,and the coding unit is terminated in advance to continue the downward segmentation process.Experimental results show that,compared with the original HTM13.0,the algorithm saves 1.3%of the encoding bit rate and saves 23.3%of the overall encoding time.3?This paper proposes an algorithm based on information entropy for early termination depth map quad-tree segmentation.The algorithm aims at determining whether the current coding unit needs to be further divided by setting a fixed threshold.Because uses the two-dimensional entropy value of the image which can reflect the richness of the image texture.Firstly,we use the rate-distortion comparison in the official algorithm to obtain a segmented and undivided coding unit dataset.Secondly,the information entropy value that distinguishes the divided and non-divided coding units to the greatest extent is obtained through data statistics,and this number is used as a threshold value to determine whether the depth map is smooth.Finally,before performing quad-tree segmentation of the depth map coding unit,calculate its two-dimensional entropy at first.If it is smaller than the information entropy threshold that obtained from the previous statistical data,it will directly terminate the subsequent rate distortion recursive calculation.Experimental results show that,compared with the original HTM13.0,the algorithm saves 36.7%of the overall coding time,and the coding bit rate only increases by 0.5%.