Research On Compression Algorithm Based On Video Codec Standard And Deep Learning

Video compression is a necessary and constantly challenging problem.A variety of video codec standards have been derived for this problem,among which HEVC is a relatively mature codec standard and brings higher efficiency to video compression on the basis of AVC.This paper proposes a deep learning-based filtering algorithm for HEVC,which aims to improve the reconstruction quality of HEVC image frames,so that at the same or less coding cost,a better quality reconstructed image is obtained through the proposed filtering algorithm to further reduce the artifacts generated by block coding.The main work of this paper is as follows:Firstly,a multi-level feature fusion residual network is proposed for the video compression configuration of the full I frame,and a multi-level feature connection residual block is used for filtering the image decoded by the HEVC with only intra prediction to improve the reconstruction quality.The multi-level feature fusion residual network preserves the global features of each level by merging the network multi-level features to the output,and combines the multi-level feature fusion residual blocks to preserve the local features to maximize the extraction of effective features.At the same time,the network structure is combined with the residual network to learn the residual of the filtered image and the original image to quickly optimize the filter.Among them for the specific situation of the video sequence of the YUV420 format,the network structure is improved,and the up-sampling network is added to fuse multiple channels as the input of the filtering network.Secondly,on the basis of including inter-frame prediction,the relationship between different HEVC ring filtering algorithms and network filtering algorithms is considered,and the out-of-loop filtering method is improved to the network filtering mode of intraring filtering,and the above network is further improved.Moreover,for the reconstructed images of multiple quantization parameters,different statistical characteristics need to separately train the network,and the HEVC coding intermediate information of the QP map is added in the network to guide the filtering operation.Finally,the feasibility of the algorithm is verified through experiments.The experimental results show that the average BD-rate of the reconstruction algorithm of the multi-level feature fusion network used for full I-frame filtering can reach-1.88%compared to the original HE VC algorithm.For the case of interframe prediction,the proposed method can achieve an average BD-rate of-2.03%compared with HEVC.