Adaptive Selection Of Reference Picture In Video Coding

The wide application of digital video technology and the rapid development of the information technology industry have created a demand for visual quality of ultra-highdefinition video.As a response to such a demand,the video coding technology has been updated and ungraded steadily.As two of the key technologies in the field of video coding,the inter-frame prediction reference technology and the rate-distortion optimization technology have always been the hotspot in the research of video coding standards.This thesis focuses on the video coding reference structure and rate-distortion optimization problem,and proposes optimization algorithms at the picture-group level,video-frame level and coding-block level respectively.The research content of this thesis includes three aspects.(1)The adaptive reference picture management mechanism for video coding and coding parameter optimization algorithm in the picture-group level.For the low-latency coding structure,after implementing the adaptive reference picture management mechanism,this thesis proposes a long-resident reference picture scheme to realize the reference relationship over the ultra-long reference distance.Meanwhile,a reference structure with cascaded picture group is designed as a feasible scheme for selecting and scheduling long-resident reference pictures,with which the quantization parameters of each frame are optimally adjusted at the GOP level.(2)Frame-level adaptive coding parameter optimization algorithm based on temporal reference rules.For the random access coding structure,by means of analyzing the preconfigured temporal reference rules of the encoder,this thesis counts the reference number of each image frame,and then designs a model to calculate the reference value of each image frame.With the reference value of the image frame,the quantization parameters and Lagrange multipliers for each frame image can be adaptively adjusted at the video frame level.(3)Block-level adaptive Lagrange multiplier optimization algorithm based on core reference frame.By analyzing the motion intensity of the core reference frame in the source video sequence,this thesis gives a model that can compute the motion-compensation information entropy of each maximum coding unit in the core reference frame.Then the classical R-D exponential model is applied to derive the optimal Lagrange multipliers at the maximum coding unit level.This algorithm applies the rate-distortion optimization technology to the maximum coding unit level,fully considers the difference between the maximum coding units caused by the characteristics of the source video itself,and adaptively optimizes the Lagrange multiplier of each coding unit.The adaptive reference image management mechanism and optimization algorithms in three different levels proposed in this thesis have been integrated and tested on HPM,the open source encoding platform of AVS3 standard,and VTM,the open source encoding platform of VVC standard.The test results show that all the three optimization algorithms can achieve good performance gains,which can ensure both objective and subjective quality of the video while improving the coding efficiency.Part of the output in this thesis has been adopted by the China Audio and Video Standards Working Group and has formed a patented technology.