Optimization On Matrix-Based Intra Prediction Algorithm In H.266/VVC

With the increasing demands of video services in people's daily life,the compression of video data has become a necessary topic in multimedia communication.Video coding is an important technique of compressing large amounts of video data into a bitstream.In order to meet the reduction of the bit rate of video transmission,the next generation video coding standard called H.266/Versatile Video Coding(VVC)emerged as the times require and has been released in 2020.Matrix-based Intra Prediction(MIP)is the first coding tool which applies neural network to coding framework.As a new technology introduced by H.266/VVC,MIP mode predicts through multiple pre-stored tables in memory,which effectively compensates for the shortcomings of traditional intra prediction coding and plays an important role in video coding tools.However,the early MIP introduced into H.266/VVC needs improvement in the pre-and post-processing,and the use of memory space in the calculation process is also redundant.In order to optimizes the above-mentioned shortcomings,during the period of participating in the formulation of H.266/VVC,this paper conducts an in-depth study of MIP and proposes some methods,in which the matrix coefficient derivation and representation algorithm has been adopted by H.266/VVC.The specific research contents and innovative work of this article are as follows:1.In order to improve the coding performance,two MIP algorithms are proposed based on reasonable hypothesis aiming at the pre-and post-processing of MIP:(1)A coding block classification optimization algorithm based on texture features is proposed.Through the comparative analysis of H.266/VVC coding block size and MIP classification,the algorithm adjusts the classification index with the size of 8 × 8 and applies the optimized coding block classification to the prediction pre-processing which improves the coding performance without affecting the computational complexity.(2)An up-sampling optimization algorithm based on bidirectional averaging is proposed.By analyzing the up-sampling process and combining with the research of other adopted upsampling methods,the algorithm applies the twice parallel up-sampling method to the postprocessing to improve the coding performance.2.In terms of complexity reduction,aiming at the core process of MIP,this paper analyzes in detail the mapping relationship and mutual transformation between the original floatingpoint matrix and the integer weight matrix of standard.Based on this,three complexity optimization schemes and a corresponding optimization algorithm are proposed.(1)An optimization scheme of matrix based on weight analysis is designed.Through the analysis of the isolated points of the weights,based on the assumption that some weights of the floating-point matrix are inverted,the method of using the opposite operation on some of the input vector is adopted to map the updated floating-point matrix to a new integer matrix.The scheme reduces the weight range without changing the coding performance and saves some memory space.(2)A simplification scheme of shift of weight table based on unified operation is designed.Through the analysis of accuracy of the weights,the values in the shift of weight table are unified to remove the table for reducing the calculation complexity,which saves a few memory overheads with little performance loss.(3)A simplification scheme of factor of offset table based on unified operation is designed.Through the analysis of the weight range,the factor of offset table is removed by a fixed value without affecting the coding performance.(4)An optimization algorithm based on retraining matrix is proposed.Based on the above three schemes,this algorithm traces back to the weight training process,designs a nonlinear function to limit the training weight range,retrains the matrix and then converts them into integer matrix to replaces the original matrix.This algorithm reduces computational complexity while gaining coding performance improvement,which saves large quantity of memory space.It has been adopted by H.266/VVC and has been integrated in the reference software of H.266/VVC.