Research On Fast Intra Coding Algorithm Based On Regional Decision Trees For Virtual Reality 360-degree Videos

ERP(Equi-Rectangular Projection)format is a commonly used projection format for 360-degree videos.The video images near two levels have horizontal stretching and distortion,which results in the performance degradation of the High Efficiency Video Coding(HEVC)encoder.In order to improve the encoder performance,decision tree models are used to optimize the CU(Coding Unit)partition process and the reference pixel filtering process of intra coding.The main work is as follows:(1)Aiming at the high coding complexity of CU partition in 360-degree video intra coding,the paper researches the optimization structure of CU partition,and uses the RD-cost of adjacent same-sized CUs and the depth correlation between CUs to construct the decision tree attributes.Different classification categories are established for different regions and different types of decision trees.Based on the decision tree models,a 360-degree video CU fast partition algorithm is designed.When encoding,the CU will extract different sample attributes according to its region,and determine whether the current CU is early-split or pruned according to the classification result of the decision trees.CU will skip the cost calculation process of some depths,and reduce the coding complexity.Experimental results show that compared with the standard algorithm of HM16.20,the proposed algorithm can reduce the encoding time by 28%,while BD-rate only increases by 0.27%.(2)Aiming at the inaccurate judgment of 360-degree video reference pixel filtering method,the paper optimizes the configuration of the reference pixel filtering method,increases the number of reference pixel sampling points,and designs different attributes and classification categories for different regions.Based on the regional decision tree models,a 360-degree video reference pixel filtering optimization algorithm is designed.During encoding,the filtering method of the reference pixels is judged according to the TU(Transform Unit)size and the classification result of the regional decision tree model to make the intra prediction process have higher prediction accuracy.Experimental results show that compared with the standard algorithm of HM16.20,the proposed algorithm saves BD-rate by 0.1% on average,and the encoding time remains basically unchanged.