Research On Intra Coding Fast Algorithm In HEVC

With the continuous development of Internet and multimedia technology,people's demands for video quality increase day by day.A large number of HD and UHD video made a great impact on computer storage space and transmission bandwidth.To ease the problem caused by the rapid growth of video data,JCT-VC developed and released HEVC,the new generation of video coding standard.Compared with H.264/AVC,the last generation video coding standard,HEVC can save half of the bit rate under the same coding quality.The advanced coding technologies not only enhance the coding efficiency,but also introduced greatly computational complexity to encoder,which make HEVC difficult to expand in practical application.Therefore,reducing the computational complexity became a research hotspot in video processing field.To solve the problem mentioned above,this research focuses on the fast algorithm for intra coding in HEVC.The mainly work of the thesis is as follows:Considering the relevance between videos' textural feature and intra prediction modes,two fast algorithms of prediction modes selection are proposed in this paper.The first method designs four directional detectors to obtain angular information,then the high likely modes pre-grouped are selected to make up mode candidate list according to the angular information.And the second algorithm allows the encoder to traverse the prediction modes adaptively according to the H cost.The prediction modes with high H cost will be excluded from the RMD and RDO modes list to reduce the coding time.Besides,since machine learning performs well on classification questions,we utilize the support vector machine to speed up the CU size decision process.First,we design three texture extracting models to extract the CUs' textural features.Then,the training samples are selected from the CPIH database and used to train the cascade SVM models offline.Finally,the trained cascade SVM models are transplanted into the encoder and decide CUs' depth early to avoid unnecessary traverses or prediction process.To validate the performance of our algorithms,all methods proposed in this thesis are implemented on HM 16.0.Experimental results show that,the fast prediction modes decision algorithms can achieve 16.0% and 16.8% average encoding time reduction,respectively,without bitrate increase.And the fast CU size decision algorithm can save 43.8% encoding time on average with only 0.24% bitrate increase while causing negligible visual performance loss.Above all,the results verify that our algorithms can reduce encoding time effectively with high coding quality.