Research On Joint Optimization Of Video Coding Reference Frames Based On Deep Learning

The quality of reference frames directly affects the video coding efficiency.Recently,Convolutional Neural Network(CNN)-based reference frame enhancement and synthesis have exhibited superior performance over traditional methods.However,state-of-the-art methods generally focus on optimizing individual coding tool,e.g.,the frame enhancement or synthesis technique is developed and deployed separately in video codecs.While in the coding loop,the two modules are correlated to each other,e.g.,the enhanced frames will be referenced in the subsequent inter frame coding,which requires a joint optimization on the them.In this thesis,we propose a CNNbased joint optimization approach for the two modules to work consistently in video encoders.Specifically,the reference frames are enhanced in the in-loop filtering stage by exploiting the spatial correlations across pixels,and the virtual frames are synthesized in the motion compensation stage by exploiting the temporal correlations across neighbouring frames.Besides,a collaborative training method is utilized to train these two CNN-based modules through simulating the reference dependencies across frames.The main innovations are as follows:1.A joint optimization on reference frame enhancement and reference frame synthesis modules is proposed to extract temporal information by reference frame synthesis in the inter-frame prediction stage and spatial information by reference frame quality enhancement in the intra-loop filtering stage;and finally the spatialtemporal information is collaboratively utilized to improve the coding efficiency.Experimental results show that the proposed method achieves averagely 10.36% BDRate coding gain over H.265/HEVC in RA configuration,which surpasses state-ofthe-art CNN-based in-loop filtering and frame synthesis methods.2.For reference frame enhancement,the Reference Enhancement Network(EnhNet)is designed and a frame-level mode selection strategy based on RateDistortion Optimization(RDO)is proposed.For reference frame synthesis,the Reference Generation Network(GenNet)is designed to adaptively compensate the CTU(Coding Tree Unit)without additional motion estimation and motion compensation processing..3.Considering that EnhNet and GenNet will work together in the coding loop,a collaborative training strategy including the progressive training for EnhNet and the cycle consistent training for GenNet is proposed to deal with the complex reference dependencies in inter-frame coding during model training.This practically avoid the data overfitting caused by iterative filtering propagated across temporal reference frames and mitigate the iterative CNN processing effects across inter frames.