Research On Fast Algorithm Of 3D-HEVC Intra Depth Map Coding

With the rapid development of digital video technology,the application of 3D video in life is becoming more and more.Such as 3D film and television,3D projector,VR.The data volume of 3D video is usually very large,because 3D video usually needs more 2D video screens as information carriers,and also needs to record a large amount of auxiliary information and distance information.This undoubtedly brings great challenges to the storage and transmission of video.In order to ensure the transmission and compression efficiency while ensuring the quality of 3D video.The Joint Collaborative Team on 3D Video Coding Extension Development(JCT-3V)established by the Video Coding Experts Group(VCEG)and Moving Picture Experts Group(MPEG)has developed a 3D high-performance video coding framework(3DHEVC)based on the 2D high-performance video coding framework(HEVC).3DHEVC uses many new technologies,such as TDOA compensation prediction,viewpoint synthesis,distortion optimization,depth map and so on,which greatly compresses the volume of video.However,the coding complexity of 3D-HEVC is high,and the long coding time hinders the promotion of 3D video.Therefore,further research and improvement are neededThis paper mainly optimizes the intra quadtree coding structure of depth map introduced by 3D-HEVC.Similar to the texture map coding inherited from hevc,the intra coding of depth map also adopts quadtree coding structure and intra mode selection,which is one of the links with high time complexity.At the same time,because of the sharp edges and large smooth areas,compared with the quadtree coding of texture map,the quadtree coding of depth map has more redundant calculation.In order to improve the efficiency of depth map intra coding,the following methods are proposed in this paper:1.Proposed an algorithm based on two-dimensional entropy and variance.Firstly,through the data statistics of two-dimensional entropy and variance of Coding Unit(CU)of a large number of intra depth maps,the decision thresholds of twodimensional entropy and variance are obtained respectively.During coding,if the value of the two-dimensional entropy of the current CU is less than the two-dimensional entropy threshold,it will not be divided.If it is greater than the threshold,it will continue to be judged by variance.If the variance value of the current CU is greater than the variance threshold,continue to divide,otherwise do not divide.Thus,the calculation is terminated in advance and the size of the CU is determined.Comparative experiments show that the proposed algorithm can reduce the computational complexity,save coding time,and the average video quality loss is within an acceptable range.2.Proposed a fast depth map segmentation algorithm based on adaptive convolutional neural network.In view of the problems that the traditional fast segmentation method of depth map needs to manually select image features,difficult to extract the correlation between multiple features,and the generalization ability of the algorithm is not strong.This algorithm proposes to use the way of deep learning to divide the CU of depth map.The size of the CU to be decided in the depth map is 64×64?32×32 and 16×16 three,which makes the network model difficult to unify.In this paper,a CNN network with adaptive CU size input is proposed.Spatial pyramid pooling(SPP)is used to fully extract the characteristics of three size CU,which solves the problem of unified input convolutional neural network of multi size CU.3.Proposed a fast depth map segmentation algorithm based on Non-local Self Attention(NLSA)convolutional neural network.There are large-scale flat areas and sharp edge areas in the depth map,which is far from the texture map.In this paper,NLSA is used to improve the adaptive convolutional neural network model,so that the proposed convolutional neural network can skip the large-scale flat areas in the depth map and pay more attention to the areas with sharp edges,so as to achieve a good prediction effect on the CU of the depth map.The comparative experiments show that the proposed algorithm saves a lot of coding time and has good coding quality.