Research On The Inter Prediction Algorithm For HEVC Based On Neural Networks

Video have become the key component of people's daily life and Internet data,since it is the fastest and most efficient data form for people to obtain information.As the Internet and self-media are thriving,massive amounts of video data are generated in the Internet every day,which brings challenges to the network bandwidth and the storage capacity of terminal devices.Therefore,video coding,which aims to reduce the amount of video data without introducing much distortion,has become one of the core technologies of many video-based application scenarios.High Efficiency Video Coding(HEVC)mainly reduces spatial redundancy,temporal redundancy and frequency redundancy by intra-frame prediction,inter-frame prediction and transform and quantization,respectively.Among these parts,inter-frame prediction is used to remove the temporal redundancy between frames,and its efficiency greatly affect the overall coding efficiency.In the inter-frame prediction,the motion estimation and motion compensation are conducted under the guidance of the reference frame,whose quality will greatly affect the efficiency of inter-frame coding.In order to improve the quality of reference frames,many researches have improved the efficiency of inter-frame coding by generating new reference frames by using video frame interpolation algorithms.However,these algorithms are less robust while encountering videos with motion blur.Motion blur commonly exists in daily shooting videos.Varying motion blur will reduce the correlation between adjacent frames.In this case,some studies attempt to use filter-based methods to improve the performance of inter coding for videos with motion blur.However,these methods have the following problems,such as requiring prior information of motion vectors and poor generalization of filter coefficients,while delivering low-efficiency interframe prediction in motion blur scenes.In order to improve the quality of reference frames during the inter-frame coding for motion-blurred videos,and overcome the limitations of the existing methods,this thesis proposes a novel inter-frame coding algorithm for motion-blurred videos,which is based on reference frame prediction network.Specifically,the main work of this thesis consists of the following three parts:(1)Establishing a motion blur dataset for the video coding scenario.(2)Designing a reference frame prediction network for motion-blurred scenes.(3)Integrating the proposed reference frame prediction network into HEVC.Compared with the reference software of HEVC(i.e.,HM16.9),the proposed algorithm achieves on average 1.21% BD-rate reduction under RA configuration.Specifically,the proposed method achieves on average 1.55% BD-rate reduction for motion-blurred videos,and also obtains 0.65% BD-rate reduction for the sequences provided by MPEG.All results have demonstrated that the proposed algorithm can improve the efficiency of the inter-frame coding for motion-blurred videos,while delivering a robust performance for general video sequences.