The Methods Of Depth Maps Processing Basded On Compressive Sensing And 3D Video Coding Optimization

Three dimensional(3D)video has been the goal of video processing and communication because it can provide more realistic and natural visual experience.Free Viewpoint Video(FVV)is the highest form of 3DTV.It can change viewpoint arbitrarily according to the location of viewer.Faced on the huge amount of data,the deep distortion of depth maps under unreliable network environment and the low quality of virtual view after decoding and other problems,the traditional video or image compression technology has been unable to meet the needs of 3D coding system.The emergence of compressive sensing theory provides a new idea.Based on the deep analysis of image/video compressive sensing,distributed video coding,free viewpoint coding and virtual viewpoint rendering theories,this paper proposed a new depth maps processing system based on Compressive Sensing,and researched on a number of key optimization methods of 3D video coding.,the main achievements and innovations include the following aspects: 1 Adaptive hierachal sampling and reconstruction for depth maps based on Block Compressive Sensing.Aiming at the problems of high computation cost,fixed sampling rate and uncleared priori information,an adaptive hierachal sampling and reconstruction for depth maps based on BCS for depth map is proposed.In terms of compressive sampling,the modle of an adaptive hierachal block sampling is designed based on multi-scale characteristic of the wavelet domain,the difference influence for rendering of the depth under various scales and the local sparseness in the measurement domain,which can measure weightedly in accordance with the sparse of coefficient block to improve the sampling efficiency.On the basis of multi-level sampling model,an adaptive reconstruction algorithm based on depth image entropy is proposed.Bilateral filtering is used as a smoothing filter in the iterative reconstruction of the algorithm to further smooth the block effect and protect the depth map.The experimental results show that the proposed algorithm has a performance improvement of 0.2 ~ 1.8dB for the quality of the synthesised viewpoint compared with the traditional algorithm.From the subjective test,the proposed algorithm can be improved in the synthesis boundary and the prevention of jitter and ghosting.2.Research on the distributed compressive sensing system for depth sequence combined spatio-temporal characteristicOn the basis of analyzing the advantages and disadvantages of the different distributed compressed sensing model,a distributed compressive sensing system for depth sequence combined spatio-temporal characteristic is designed using the correlation of the spatio-temporal domain.Based on the model of distributed compressive sensing recovery of depth map using multihypothesis predictions,we introduce a new model of linear TV norm as the constraint of spatial dimension to decrease distortion in the process of depth reconstruction.Experimental results show that the proposed method is about 0.2 ~ 0.6dB higher than that of other models.In addition,this paper discusses the different prediction modes of the system,and verifies the influence of different modes on the virtual viewpoint synthesis through experiments.3.The methods of free view viewpoint coding system optimization(1)Texture Video Coding System OptimizationFor texture video,an efficient HEVC coding framework is used for encoding.Aiming at the problem of high complexity of HEVC coding,a fast coding unit size decision in HEVC intra coding is proposed in this paper.Fast coding unit size decision in HEVC intra coding.Three aspects are a fast coding unit select algorithm based on the content of image,a rate distortion optimization method based on the training frame and a fast CU choice algorithm based on residual distribution.Experiments show that the proposed algorithm has a 50.4% decrease in coding speed compared with the reference software HM14.0,and only 0.08 dB PSNR loss and 1.37% BDBR rise.(2)Research on Side information prediction of Distributed Compressive Sensing for Depth SequenceThe prediction of side information is an important component of distributed compressive sensing system for depth maps.The quality of prediction affects directly the performance of reconstruction.According to the characteristics of the depth maps,this paper presents a side information prediction algorithm based on multi-scale boundary protection for depth maps.Firstly,the multi-scale wavelet transform is used to construct the depth maps pyramid model,and construct matching criterion of the boundary protection.Then,smooth movement and purification are process step by step form the bottom of the pyramid.In order to further enhance the algorithm performance,the depth blocks are divided into continuous and discontinuous blocks.The initial motion vectors are updated in different ways.Experiments show that the side information prediction method proposed in this paper has an average 7.32% depth image error rate drop and 0.53% lower edge error rate than other algorithms.The proposed algorithm has a very good effect both from the depth distortion and the subjective quality of the synthesised viewpoint.