Optimization Of Intra Cross-Component Chroma Prediction Algorithm Based On VVC

Inter-component correlation is one of the typical redundancies in video data.In the Versatile Video Coding(VVC)standard,the Cross-Component Linear Model(CCLM)is introduced to reduce the inter-component correlation,which has achieved significant results.However,The CCLM model is limited in its ability to express the correlation between luma and chroma,and is not effective in predicting complex coding blocks.A study is conducted to address this problem and cross-component prediction algorithms based on the coloring principle and asymmetric convolutional networks are designed to improve the inter-component prediction performance.The main contents and innovations are as follows:(1)The inter-component correlation is studied from the perspective of coloring principle.Through statistical analysis of the reference video sequences,it is found that the Y,U and V components always show similar coloring patterns.Based on this similarity,a cross-component chroma prediction algorithm based on the coloring principle is proposed to predict the U and V components with the coloring pattern of Y component.The problems of selecting the position and number of coloring samples are investigated,and the optimal selection scheme for two sets of coloring samples is designed from the perspective of balancing the complexity and performance of the algorithm.Meanwhile,the optimization problem of coloring coefficients is investigated with luma information for a lightweight description of the coloring relations and the parameter is optimized with a statistical approach.In addition,mode encoding is defined for the two designed coloring principle-based chroma prediction modes.Experimental results show that 0.03%,1.16% and 1.17% BD-rate reduction in the Y,U and V components is obtained in the proposed algorithm,respectively,with negligible increase in coding and decoding time compared to the reference model VTM12.1 in the All Intra(AI)configuration.(2)To address the problem of poor performance for complex coding block prediction with CCLM model,the algorithm of cross-component chroma prediction based on deep learning is investigated and a cross-component chroma prediction algorithm based on asymmetric convolutional network is proposed.To improve the utilization of spatial correlation of adjacent samples,an asymmetric convolution module is introduced in the algorithm;To further extract the spatial correlation of the current luma block,an improved Asymmetric Convolution Block(IACB)is designed in the algorithm,which takes advantage of asymmetric convolution to enhance the central skeleton of the square convolution kernel;In addition,to make full use of the features extracted from different layers of the network,Feature Reuse Block(FRB)is introduced in the algorithm.The experimental results show that 0.62%,3.51% and2.67% BD-rate reduction in the Y,U and V components is obtained in the proposed algorithm,respectively,under the AI configuration compared to the reference model VTM12.1 under the constraint.