Research On Fast Intra-frame Coding Of Depth Images In 3D-HEVC

In recent years,apart from the clarity and fluency of video,people begin to pursue better video watching experience and hope to watch three-dimensionality videos,which push the development of the 3D-HEVC encoding technology.The Coding Tree Unit(CTU)of 3DHEVC is the same as that of HEVC,and adopts a recursive hierarchical structure based on a quadtree.At the same time,in order to preserve the high-frequency information of sharp edges,some new coding modes are introduced for depth images.The complex quad-tree structure and a large number of predictive coding modes of the coding unit(Coding Unit,CU)dramatically increase the computational complexity of the depth image.This paper aims to reduce the complexity of depth image coding,and based on the deep learning method,two fast depth map intra-frame coding algorithms are proposed.The main contents are described as:(1)Aiming at the complex quadtree partition structure of the depth map in 3D-HEVC,a CU partition prediction network(Partition Decision Network,PDNet),based on a multi-scale convolutional neural network,is proposed to achieve fast CU partition decision.The recursive partition structure of the quadtree enables the correlation between the partition labels of different depths of CU.This paper projects this correlation into the neural network,specifically,the network employs deconvolution to aggregate the global and local features of the coding unit,then outputs all partition prediction results of a CTU at one time.According to the coding block characteristics and prediction probability distribution of different QP and different depths,a prediction probability threshold that varies with QP and depth is designed.The experimental results show that,with only 0.50% increase in BDBR,the proposed PDNet saves 54.55% of coding time.(2)Aiming at the complicated intra-frame prediction coding mode of depth map in 3DHEVC,this paper classifies the prediction modes into DIS mode and other Intra modes,on the basis of the rate-distortion performance analysis of DIS mode.Through analyzing the feature commonality of DIS mode and partition label,based on the feature extraction module of PDNet,this paper proposes a partition-and mode-decision network(Partition-and Mode-Decision Network,PMDNet).In the end,with an increase of 0.89% in BDBR,the proposed algorithm saves 59.12% of coding time.