Research And Application Of AVS3 In-Loop Filter Technology Based On Neural Network

With the vigorous development of information technology,emerging multimedia applications are gradually becoming popular,and video as a key information carrier has a wide range of applications.Audio Video Coding Standard(AVS3)is a new generation of video coding standard developed by China,in which in-loop filter technology is used to reduce distortion in the coding process,which can effectively improve video quality and further enhance compression performance.In this thesis,we thoroughly study the in-loop filter technique for intra-frame coding in AVS3,and propose an optimization approach for the neural network-based in-loop filter algorithm that considers both the coding block boundary and internal aspects.For the optimization of coded block boundaries,a partition-aware inloop filter algorithm based on depth separable convolution is proposed in this thesis.This algorithm utilizes the block partition information during the encoding process to assist in the in-loop filter procedure.It represents the encoding block information using a local mean-based method,introduces deep separable convolution to extract block boundary features,and further combines these features with decoded frames suffering from distortion.This improves the reconstructed frame quality effectively,and thus enhances the encoding performance of AVS3.This algorithm divides the model training based on quantitative parameters(QP),and controls frame-level and encoding-unit-level switches through the rate distortion optimization process.Finally,a partition-aware model is used to filter the luma component,which is used as a guide for chroma filtering instead of traditional in-loop filter techniques.Experimental results show that compared with traditional in-loop filter algorithms,the proposed blockpartition-aware in-loop filter algorithm can save 5.50%,5.61%,and 8.61%of BD-Rate(Bj?ntegaard Delta Bitrates)for Y,U,and V components,respectively,in an all intra configuration.Compared to existing neural network-based in-loop filter,it can save 0.36%,1.38%,and 0.62%of BDRate for Y,U,and V components,respectively.For the optimization within the encoding block,this thesis proposes a multi-level collaborative attention in-loop filter algorithm.This algorithm utilizes attention mechanisms to adaptively generate residual images,and introduces efficient and flexible attention modules and residual feature aggregation modules.The attention module allows the model to pay more attention to salient features to better capture visual structures.The spatial attention aims to enhance the feature expression of key areas,while the coordinate attention aims to simulate inter-channel relationships and capture long-term dependencies with precise position information.The residual feature aggregation module sends the features of local residual blocks directly to the end of the model to fully utilize the locally refined features generated by the attention module.In addition,the model uses prior information QP as an auxiliary feature input,making a single model suitable for handling encoded videos with multiple QPs,and the effectiveness of the method is verified through experiments.The luminance-guided filtering is also applied to chroma components.Experimental results show that the proposed multi-level collaborative attention loop filtering algorithm can save an average BD-Rate of 5.58%,6.16%,and 8.82%for Y,U,and V components,respectively,compared to traditional in-loop filter algorithms in an all intra configuration.Compared to existing neural network-based in-loop filter algorithms,it can save an average BD-Rate of 0.40%,1.90%,and 0.82%for Y,U,and V components,respectively.