Research On Fast Transcoding Algorithm From H.264 To HEVC Based On Analyses Of Coding Unit Features

As one of the most important standard in the area of video compression,H.264 occupied a large portion of market shares in the last decade due to its superiority in the compression rate,flexibility and network transmission performance.However,as the prevalence of High Definition(HD)and Ultra High Definition(UHD)videos,H.264 is unable to meet the demand for video compression rate because of the limitation of its coding structure.Consequently,HEVC has been taking place of H.264 and becoming the priority of various platforms.Therefore,it is of great market value to investigate the transcoding algorithm from H.264 to HEVC under such a background.Traditionally,the cascaded transcoder fails to meet the requirement of the real-time communication system because of the huge computational complexity of HEVC re-encoding process,which is now the main problem preventing it from being applied in markets.Cosindering this problem,we make an elaborate investigation and propose two algorithms to accelerate the process of inter-frame coding unit(CU)depth decision and prediction unit(PU)mode decision,which are the most time-consuming processes in the HEVC re-encoding process.In terms of the inter-frame CU coding depth decisions process,we firstly make a detailed statistics analysis of the mapping relationship between the best CU depths and the features extracted from H.264 bitstreams.Then,we propose a three-output classification model to predict whether the current CU should be split or not.In order to increase the prediction accuracy of this model,we also propose a specific building method of feature vector and the designing method of three-output classifier.During the reencoding process,this classification model can assist the transcoder to adaptively bypass or stop the unnecessary checking procedures for certain sizes of CUs.Experiment results present that this algorithm can help to save about 44.06% of transcoding time with 1.21% of BDBR degradation under LowDelay_P configuration.In terms of inter-frame PU mode selection processes,a hierarchy stopping scheme for PU mode selection is proposed.This scheme contains three sections,including early Skip mode decision method,early Merge and 2Nx2N(M&S)modes decision method and early intra mode exclusion method.With the help of the relevant decoding information,these methods are able to early screen out the modes with high occupation rate and early exclude the modes with little occupation rate.Experiment results present that this algorithm can help to save about 35.14% of transcoding time with 1.47% of BDBR degradation under LowDelay_P configuration.Lastly,we evaluate the performance of overall algorithm including the fast CU depth decision method and the hierarchy stopping scheme for PU mode selection,and the experiment result shows that the overall algorithm can achieve 55.30% of time reduction with 2.08% of BDBR degradation.In this paper,the mapping relationships among the best CU sizes and PU mode and the dcoding information are analyzed,which are then combined with machine learning method to reduce the unnecessary checking procedures of CU sizes and PU modes.Extensive experiment results prove that the two algorithms are of high efficiency in accelerating the transcoding speed,which is beneficial for the research of real-time transcodings.