Deep Learning Based In-loop Filtering In Video Coding

In the era of information,video is an important information medium.The amount of video data transmitted is growing exponentially.People's requirements for video quality and video resolution are also increasing.This brings new opportunities and challenges to video coding technology.Compared with the previous generation video coding standard H.264/AVC(Advanced Video Coding),the currently commonly used video coding standard HEVC(High Efficiency Video Coding)improves the coding efficiency by about50%,but it still cannot meet the needs of video applications.In block-based video coding,each coding unit is independent of each other and has different coding parameters,so that the coded video has block effect.In addition,the loss of high frequency components during encoding causes ringing effects.Blocking,ringing,and coding blur are collectively referred to as compression artifacts.Since the reconstructed frame is used as a reference frame,these compression artifacts adversely affect the subjective quality of the video and the coding performance of subsequent frames.In HEVC,H.264/AVC video coding standards,methods such as deblocking filter and sample point adaptive compensation are used to improve the problem of compression artifacts.But these methods are manually designed and difficult to adapt to complex video content.With the development of deep learning,filters based on convolutional neural networks can bring better results than traditional filters.However,these existing filtering methods based on deep learning have shortcomings,and they do not analyze the filtering process in combination with the characteristics of video coding.Aiming at the characteristics of video coding compression artifacts and using the advantages of convolutional neural networks,this thesis proposes two algorithms to reduce compression artifacts and improve video coding efficiency.The specific work is as follows:1.Intra-frame enhancement method based on feature mixing: This method captures high-level semantic information of video reconstruction frames through a semantic extraction network.The video structure information is obtained through a U-Net-like structure-semantic feature extraction module with pooling operation.Additionally,lowlevel texture distortions in video coding are learned through a texture feature extraction module implemented by multiple stacked convolutional layers.Experiments show that the method brings 11.3% BD-rate savings in All-intra mode compared to the HEVC coding standard.2.Video filtering method based on the fusion of global representation and local coding distortion: This method improves compression artifacts from the process of image denoising and coding distortion recovery,and designs a fusion network with three branches.One branch focuses on global contextual feature extraction for the denoising process.One branch focuses on high-level local feature extraction for the coding distortion recovery process.Another branch focuses on the extraction of basic semantic features and supports spatially accurate mapping.The three branches are merged together by an attention-based fusion method to restore the original video.Experiments show that this method achieves BD-rate savings of 13.5%,11.3%,and 11.7% in All-intra,Lowdelay,and Random-access coding modes,respectively,compared to HEVC.