The Research And Optimization On The Technology Of Depth-Image-Based Rendering

The introduction of depth image in 3D video coding and view synthesis technology marks the true change of 3D TV from traditional 3D-TV to FTV,because it allows users to change the viewpoint as desired,thus giving people a more immersive threedimensional feeling.Compared with the traditional 3D-TV,Depth-Image-Based Rendering(DIBR)does not need to encode and transmit a large amount of multi-view video data,but only needs to acquire the texture and depth of a limited number of viewpoints.Other plurality of intermediate views can be generated by using warping and merging techniques in view synthesis system under the condition that the camera parameters are given.Due to the fast speed and high synthesis quality,this technology has been adopted by many 3D video codec technology standards such as 3D-HEVC(Three Dimensional High Efficiency Video Coding)and domestic AVS2-P2-3D.However,this technique still produces a large number of holes in the synthesized views,and the synthesis process relies heavily on the accuracy of the depth image.Inappropriate filling methods will cause blocky noise in the synthesized views,and the inaccurate depth image will cause foreground corrosion or background noise at the boundary of the object in the synthesized views,which heavily affects the viewer's viewing experience.Therefore,how to fill the holes more accurately and how to correct the depth image to make the synthesized views better,have become the main topic of this thesis.To solve the above two problems raised,the research and main optimization work of this paper are as follows.The optimization is mainly carried out in the VSS of the domestic AVS2-P2-3D reference software RFD:1.Improvement of depth image boundary detection and hole expansion technology: VSS software uses double-valued boundary to mark depth image boundaries,and the boundary pixels do not participate in mapping.In this paper,it is found that using singlevalue to mark is better.The reference software uses hole expansion technology to remove depth boundary noise,which causes the loss of a large number of original pixels,resulting in a decrease in the quality of the synthesized views.This thesis proposes to turn off this technique.2.Hole filling using surrounding and occluded pixels in view synthesis: This technique improves the hole filling technique in VSS of the AVS2-3D reference software RFD6.01.According to the generation mechanism of the hole,the surrounding background pixels of some holes are completely occluded by the foreground pixels due to large inter-view shift,which causes no background pixels existing on the left and right sides of the hole.This thesis proposes a new filling scheme for such holes: using the surrounding background pixels of the holes and occluded background pixels during the mapping process as the filling template.During the experiment,it is found that the depth images given by the current test sequences are absolutely inaccurate.Therefore,this paper introduces the method of mapping the left and right reference views as the standard for determining the accuracy of the current depth image.If the depth map is accurate,then the technique proposed is applied;if not,then the filling technology remains unchanged.3.Depth image filtering technology based on MRF: Experiments have demonstrated heavy noise at the boundary of the depth image in the test sequences,which affects the quality of synthesized views.In this thesis,a depth image filtering method based on MRF(Markov Random Field,MRF)is proposed.The MRF model is built by using the segmentation blocks of Mean shift and SLIC(Simple Linear Iterative Cluster,SLIC)to filter the depth image.The experimental results demonstrate that the quality of synthesized view is significantly improved by using the filtered depth image for view synthesis.