Research On Cross Component Prediction Algorithm In AV1

Nowadays,with the continuous development of video services such as web conferences,video chat and online live broadcasts,the trend of video consumption is accompanied by rapid growth.While the uncompressed storage capacity of video is huge,it is not conducive to storage and transmission.In order to pursure higher compression efficiency and lower application cost,the video coding field has been conducting in-depth research on video coding standard.The AOM is committed to developing an open-source and royalty-free codec called AV1,which is based on VP9 developed by Google.AV1 aims to achieve realtime high-quality video transmission under different bandwidth conditions with reasonable complexity.The eventual realization of AV1 is to achieve almost 30% performance gain compared with VP9 and H.265/HEVC.Moreover,AV1 has extended each module of the coding process compared with VP9.Due to the existence of spatial redundancy in a frame,intra prediction is a key part of the coding framework,which can use spatial correlation to improve coding performance.In intra prediction technology,AV1 not only expands the types of directional intra prediction modes,but also adds new intra prediction coding tools.The main research object of this paper is CfL coding technology,CfL uses correlation between components to simulate the chroma prediction as a linear function of the reconstructed luma,which can effectively improve the coding efficiency.This paper conducts an in-depth study of CfL,aims to propose improved algorithms for CfL to further improve its coding performance,and then explore the optimization approach to find the well-balance between the coding performance and coding complexity.The specific research content includes three parts:1.CfL is an intra prediction coding tool for chroma components introduced by AV1,it uses the relevant information of the chroma average of the adjacent reference regions to construct a linear model,and simulates the chroma prediction as a linear function of luma to realize the prediction from the luma component to the chroma component.Although CfL brings effective coding gain,due to the diversity of correlations between the current coding block and its adjacent regions,the parameters of linear model may not be accurate enough.In AV1,luma is coded before chroma,therefore the luma reconstruction value distribution information of the coding block and adjacent regions can be referred.On this basis,this paper proposes a CfL optimization algorithm that performs weighted calculation based on the luma reconstruction value distribution.The proposed method constructs a virtual chroma block at the pixel level based on the luma correlation,the average of the virtual chroma block is used and more accurate linear model parameters can be derived.Experimental results show that the CfL optimization algorithm has a significant improvement in coding performance under All Intra and Random Access configurations.2.Based on the analysis of the correlation between the luma reconstruction value distribution between the current coding block and adjacent regions,this paper introduces a minimum distance CfL optimization algorithm based on the luma reconstruction value distribution,which is a further uses the luma correlation of the reference pixel with the closest spatial distance during the construction of virtual chroma block.Compared with the optimization algorithm mentioned before,this algorithm can reduce the number of multiplication and division operations.Experimental results show that the proposed optimization algorithm brings effective coding gain.3.Both of the two CfL optimization algorithms based on the distribution of luma reconstruction value can obtain valid coding gains,considering the complexity of the algorithm in hardware and software implementation,this paper finally proposes a simplified method of the CfL optimization algorithm.The simplified algorithm stores the reconstructed value information of the reference pixels in the adjacent regions,the multiplication operation in the implementation process is completely avoided and only one division operation is required for each pixel.Experimental results show that the simplified algorithm can effectively reduce the time complexity caused by the two CfL optimization algorithms while bringing performance gains.