| VP9 along with H.265/HEVC belongs to the latest generation of video compression standard and express the same performance in video quality and bitrate.On account of the open format and free of license fee,VP9 is getting more and more support and promotion in the field of high-definition(HD)and ultra-high-definition(UHD).Compared with its predecessor VP8,it can achieve a bitrate reduction about 50%,while keeping the same perceptual video quality.Although the compression efficiency is improved significantly,enormous computational complexity is introduced due to the employment of a series of new coding techniques,which hinder HD/UHD videos in real-time applications.To address this problem,this paper optimize the coding complexity of Super Block(SB)partition decision,the most time consuming part of encoder and realize the complexity scalability control to meet the variable need of different application scenarios.For the purpose of optimizing the high coding complexity of Coding Unit(CU)partition decision and mode selection,we firstly analyze the key influence factors of Quad-tree CU depth decision process and make it possible for quick forecast,based on the difference between the features of split and nonsplit mode.Differ from the traditional classification algorithm,this paper adopt the offline-trained BP neural network as the classifier,whose inputs is the value of 36 SADs as the presentation of temporal and spatial features,and outputs is the partition modes.Besides that,the encoding performance loss is fully considered in offline BP neural network training.To simplify the structure of the proposed model and promoting the performance of the classifiers,we model the complicated Quad-tree CU depth decision proces as a three-level of hierarchical binary decision problem,that is to say,getting different classifiers for different QP and block size.By combining the results of all of the classifiers,we could obtain the simplified Quad-tree structure.And the proposed algorithm only needs to deal with the simplified Quad-tree structure,which results in avoiding checking most of CU modes and getting encoding efficiency improved significantly.For the purpose of controlling the complexity scalability by one step further,we obtain the confidence of the classifier firstly and then analyze the consistency between the confidence of output mode and forecast accuracy.Considering the fact that existing a high level of consistency between them,a threshold is introduced to select the mode with bigger confidence and exclude the mode with less one.By changing the threshold value of the classifier each layer,the candidate modes of Super Block can be reduced adaptively.To control the whole complexity accurately,a candidate of threshold is built in the order of complexity control intensity.And set different threshold for each frame dynamically to control the complexity of each frame in the range of an interval.The results show that the proposed algorithm can control the encoding complexity in a big range,whose upper limit equals to the original encoding algorithm,without complexity optimization and lower limit maximize the complexity optimization with the worst RD performance loss.It reduces 40% the encoding time under the identical perceptual video quality at the same time. |
