Research On 3D-HEVC Fast Coding For Depth Maps

With the demand for higher resolution and stronger stereoscopic 3D video,multiview video information has dramatically increased the amount of data stored and transmitted.Stereoscopic video coding meets challenges and hinders the development of stereoscopic video.Faced with this problem,in order to improve the encoding performance of stereoscopic video,the Joint Group of Stereoscopic Video Coding has released a new generation of stereoscopic video coding standard 3D-HEVC,which is based on the new generation of high efficiency video coding standard HEVC(High Efficiency Video Coding).The 3D-HEVC standard follows the hybrid coding framework of the HEVC standard and adopts the multi-view video plus depth(MVD)video format,which includes multi-view texture video and corresponding depth video.3D-HEVC uses the time domain,the spatial domain and the correlation between the various viewpoints are encoded,so that the 3D video has a high compression efficiency.However,while the coding efficiency increases,the computational complexity increases dramatically,which severely limits the practical application of 3DHEVC.Therefore,it is of great significance to study how to keep the coding efficiency unchanged and reduce the coding complexity.This paper analyzes and introduces the 3D-HEVC coding framework and key coding technologies,and statistically analyzes the 3D-HEVC coding complexity.Finally,it proposes the following series of optimizations for the intra-frame prediction process with higher complexity.This paper proposes two fast intra-coding methods,the specific contents are as follows:First,for the feature that the depth map contains a large area flat area and a small part of the edge area combined with the texture type,a fast intra prediction based on bilayer texture judgment fast algorithm.The algorithm performs bi-layer texture judgment on the coding block,that is,whether the coding block is smooth.It can terminate the coding block division in advance when coding the smooth area,and skip the DMMs mode;it can skip the current size coding when coding the non-smooth area.The process is directly divided into smaller blocks for encoding.The algorithm first uses the difference matrix to calculate the gradient sum of the encoding block to express the texture complexity,and then compares the current texture complexity and sub-block texture complexity with the threshold to perform different encoding operations.The experimental results show that,compared with the original 3D-HEVC algorithm,the average bit rate of the synthesized view is only increased by 0.38%,and the average encoding time is reduced by 44.8%.Secondly,in view of the close correlation between the coding block segmentation results and the video content,an intra-prediction algorithm for rapid prediction of CU size based on ATS_CNN(Adaptive Terminated Stratification-Convolutional Neural Network)network is proposed.The ATS_CNN network is an adaptive termination layered convolutional neural network.It automatically extracts features from the video content to learn the CU division results,inputs the original CTU into the prediction model,and outputs the CU division results in a structured manner to realize the use of deep learning Reduce 3D-HEVC coding complexity.This algorithm can reduce the depth coding time by 58%,and the quality of the synthesis viewpoint remains basically unchanged.