Fast Block Partitioning Algorithm For HEVC Intra Prediction Based On Machine Learning

Compared with the Advanced Video Coding(AVC)standard,the High Efficiency Video Coding(HEVC)standard greatly improves the coding efficiency.The high compression performance of HEVC is due to the application of new techniques.However,the introduction of new techniques has increased the computational complexity of the encoder,which also hinders the application of the HEVC standard.Thus,it is very important to focus on the HEVC fast coding algorithm.This paper mainly studies HEVC intra prediction fast block partitioning algorithm based on machine learning.The main works of this paper are as follows:(1)A fast block partitioning algorithm based on Support Vector Machine(SVM)for HEVC intra coding is proposed in this paper.The innovation is to use nine characteristic vectors,including image variance,edge direction,gradient direction,quantization step and Coding Unit(CU)size to improve the classification accuracy of the SVM model.Firstly,the characteristic values are extracted from each CU block as input data in the offline phase,and the partition of the CU block is extracted as the label to train the SVM offline model.The SVM model uses the Radical Basis Function(RBF)as the kernel function,and the penalty parameter C and the RBF kernel function parameter ? are determined by the grid search method.Then the characteristic values extracted by the current CU are input into the offline SVM model in the online phase,and the offline model predicts the label of the current CU.Finally,the label of the current C U is used as the judgment whether the current C U continues to be partitioning or not.The experimental results show that the proposed algorithm reduces the computational complexity of HM13.0 to 50.5% in encoding time with 0.92% increase in the BD-Rate(Bj?ntegaard Delta Bit Rate).Compared with the classical intra prediction fast block partitioning algorithm,the proposed algorithm can save 30.1% in encoding time,and BD-Rate is also reduced by 0.87%.Compared with the literature using the SVM algorithm,the proposed algorithm can save 11.6% in encoding time with 0.13% of BD-Rate gain.(2)A fast block partitioning algorithm based on Bayesian decision for HEVC intra coding is proposed in this paper.The innovation is to remove the offline training phase of the traditional Bayesian algorithm and improve the classification accuracy of Bayesian classifier by using eight characteristic vectors,including image variance,edge direction,gradient direction and CU size.Firstly,the video frame is divided into the fas t partitioning frame and the online learning frame by using scene change detection method which is based on average grey difference.Secondly,for the video frame in the scene changes and the online learning frame,standard block partitioning is performed,and the characteristic values of the C U are extracted to establish a Gaussian mixture model.The model determines the specific parameters of the model according to the EM(Exception Maximization)algorithm,and initializes the parameter values by the K-Means algorithm.For the fast partitioning frame,the characteristic values of the current CU are input into the Gaussian mixture model,and the minimum error rate Bayesian criterion is used to judge whether the c urrent CU continues to be partitioning or not.The experimental results show that the proposed algorithm reduces the computational complexity of HM13.0 to 52.2% in encoding time with 0.99% increase in the BD-Rate.Compared with the classical intra prediction fast block partitioning algorithm,the proposed algorithm can save 30.6% in encoding time,and BD-Rate is also reduced by 0.82%.Compared with the literature which is also based on Bayesian decision rule,the proposed algorithm can save 19.3% in encoding time,while the BD-Rate only increases 0.09%.