Research On Deep Learning-Based Fast Algorithm For Video Interframe Coding

With the increasing demand for video definition,ultra-high-definition videos are gradually appearing in daily life.The massive amount of video data brings great challenges to the storage and transmission of videos.In order to effectively alleviate the contradiction between the huge amount of video data and the limited storage ability and transmission bandwidth,the international video coding standardization organization has proposed the Versatile Video Coding(VVC)standard,which aims to efficiently compress video data under the premise of retaining video quality.A variety of coding tools were introduced in VVC to improve the coding performance while greatly increasing coding complexity,which limits the application of VVC in real-time transmission scenarios such as video conferences and live broadcasting.In recent years,deep learning technology has made great progress in video analysis and processing.In order to reduce the coding complexity of VVC,this thesis uses deep learning technology to carry out the research on deep learning-based fast algorithm for video interframe coding.In order to improve the efficiency of video coding,the quad tree plus multi-type tree partition structure has been adopted by VVC,which makes the coding unit adapt to different video contents flexibly.However,the complex partition structure significantly increases the iteration times of rate-distortion optimization,which introduced extremely high computational complexity.For this reason,a deep learningbased coding unit partition early termination algorithm is proposed in this thesis.With the powerful representation capability of deep learning,the partition of the coding unit is early terminated,thereby effectively reducing the coding complexity of VVC.First,the texture and motion information of the coding unit is obtained,and its correlation with the final partition decision is analyzed.Then,an early termination decision network for the partition of coding unit is proposed,and the coding unit early termination decision is made by effectively fusing the texture and motion information of the coding unit.Finally,the proposed algorithm is integrated into VVC reference software VTM.Experimental results show the proposed algorithm can reduce the encoding time by 24.83%under the premise that the bit rate only increases by 2.52%.Inter prediction modes with high computational complexity are introduced in VVC to improve coding performance,such as affine motion compensation.In order to further reduce the computational complexity of VVC,a deep learning-based inter prediction mode early decision algorithm is proposed in this thesis.First,a Merge mode early decision network is designed,and the Merge mode early decision is made by analyzing the Merge mode prediction residual and coding context information of the coding unit.Then,a lightweight affine Advanced Motion Vector Prediction(AMVP)mode early decision network is designed to analyze the coding context information of the coding unit,and the unnecessary prediction process of the affine AMVP mode is skipped.Finally,the two networks are integrated into VTM to realize the deep learning-based early decision of inter prediction modes.Experimental results demonstrate that the proposed algorithm saves 19.98%of encoding time with the bit rate only increasing by 1.25%.