Research On Fast Algorithm Of Video Coding Based On AVS3

Video is a very important information medium in today's society.Especially in the past few years,with the rapid development of display,network,storage,and computing equipment technology,people's demand for ultra-high-definition,high dynamic range,high frame rate,multi-dimensional,real-time video is increasing,which brings greater opportunities and challenges to video coding technology.In this context,the AVS working group takes the lead in organizing the development of a new generation of audio and video coding standard AVS3.Currently,the first phase has been formulated and released,and the second phase of enhancement is also coming to an end.AVS3 integrates many technologies that improve coding performance and computational complexity at the same time,which leads to a significant increase in coding time.In order to reduce the impact of new technologies on real-time performance,this paper has conducted a lot of research on the selection of inter-frame and intra-frame prediction modes and the block partition of LCU in AVS3.The specific work is as follows:1.A fast algorithm for LCU-level prediction mode decision based on historical information is proposed.First,a visualization platform is built for the AVS3 encoder,which can quickly generate a large number of visual images with high reference value from the encoding results to analyze the temporal and spatial correlation between the prediction mode and the image content.By using the visualization platform,five universal LCU-level image content types and historical type judgment references are extracted and defined.Then,through mathematical statistics of the coding results of typical test sequences,it is proved that there is a strong correlation between the historical type judgment references and the image content types.The historical type judgment references can be used as an important basis for determining the image content types,and some prediction modes that can become the best with a lower probability can be skipped to achieve the purpose of reducing the amount of calculation.Later,the algorithm model was improved from two aspects to reduce performance loss and time complexity.On one hand,the historical type decision reference is divided into large and small pieces of historical type decision reference based on the CU size,which increases the accuracy of prediction.On the other hand,a fast algorithm is designed for special test sequences with low time correlation.The improved fast algorithm finally achieved good results.In LDP mode,the average encoding time was reduced by 17.67%,and the BD-rate of the luminance signal component was increased by 0.61%on average,which has certain reference value.2.This paper studies the block partition structure of AVS3,and proposes a fast block partition algorithm based on historical information.First,the maximum depth reference and minimum depth reference are customized,and the time correlation experiment is done with the typical test sequence of the actual partition depth.The analysis results found that the success percentages of using the maximum depth reference and the minimum depth reference prediction are 92.3%and 95.2%,and the average depth deviation of the prediction failure was only 1.07 and 1.01.Then,the problem of inaccurate prediction of the maximum depth reference for the LCU with violent exercise is analyzed,and the definition of the maximum depth reference is improved for the second time,so that the prediction success percentage is increased to 98.5%.Finally,the fast algorithm is applied to the reference code.In the LDP mode,the coding time is saved by 8.12%and BD-rate is lost by 0.26% on average.