Research On Video Quality Enhancement Algorithm For VVC Coding

In video coding,compression artifacts are one of the main factors affecting the reconstructed video quality.Existing video encoders mainly use in-loop filter to reduce compression artifacts.At present,the in-loop filter in VVC(Versatile Video Coding)mainly uses single-frame video data for filtering and reconstruction,which does not make full use of the continuity of content between frames of video sequences,so the video quality improvement is limited.To optimize this problem,the deep learning-based multi-reference frame in-loop filter uses multi-frame information for filtering and reconstruction,which enhances the video quality.To optimize the current deep learning-based multi-reference frame in-loop filter,a Transformer-based multi-reference frame fusion quality enhancement algorithm and a high-quality reference frame selection algorithm to enhance the quality of video in VVC is proposed in this thesis.The main contents and innovations of this paper are as follows.(1)In the current deep learning-based multi-reference frame in-loop filter,the feature maps of each reference frame are concatenated using the same weights.This method has two drawbacks.Firstly,the network model cannot focus on the main regions,resulting in limited reconstruction quality enhancement.Secondly,the feature maps of low-quality reference frames are concatenated using the same weights,which will affect the reconstruction quality.To solve the above problems,a Transformer-based multi-reference frame fusion quality enhancement algorithm is proposed.Multiple video frames near the current frame are selected as reference frames,and after compensating the reference frames by the motion compensation network,the feature maps of the reference frames are extracted using dense units.Followed by this,the Transformer is used to weight the extracted feature maps for fusion.In this way,the reconstruction quality is improved.The experimental results show that,in the random access configuration,the proposed method reduces the BD-BR by 6.81% and increases the BD-PSNR by 0.166 d B compared with the VTM6.0 reference model.(2)It is found that the number of reference frames and the correlation between the reference frames and the current frame directly has impact on the efficiency of the quality enhancement.In this thesis,the selection range of reference frames is expanded and a high-quality reference frame selection algorithm based on the content correlation of frames is proposed.Firstly,the increment of peak signal-to-noise ratio between coded frames is used to obtain candidate high-quality reference frames with higher quality than the current frame.Secondly,the structural similarity and the content similarity based on the difference hash are used to select high-quality reference frames from the candidate high-quality reference frames.The high-quality reference frames can be inputted to the first proposed algorithm to further improve the performance of the network.The two algorithms proposed in this paper are combined for experimental validation.The experimental results show that the BD-BR of our algorithm in this thesis is reduced by 7.34% and the BD-PSNR is increased by 0.213 d B,compared with the VTM6.0 reference model.