Research On Intra CU Partition Algorithm Based On Convolutional Neural Network

Vision is the most important way for humans to perceive the world.Human science and technology are constantly seeking to provide a better visual experience.With the rapid development of network and hardware technology,the amount of video data is also increasing day by day.In order to be able to effectively compress and transmit data,video coding technology has become the key.In response,JCT-VC released High Efficiency Video Coding(HEVC).HEVC has high coding efficiency,but its coding complexity is also very high,which seriously affects its wide application.Therefore,it is of great significance to increase the encoding speed without affecting the video quality.Convolutional neural network is a current research hotspot and has good predictive ability.In this regard,this paper proposes an intra CU partition algorithm based on convolutional neural network.First,this paper proposes an intra CU(Coding Unit)partition algorithm based on multi-level prediction.Separately train a convolutional neural network(CNN)classification model for CUs of different sizes,and use this model to calculate the current CU division probability.For CUs with a large termination probability,you can directly terminate;for CUs with a small termination probability,continue to divide;for other coding units,we combine the "skewness kurtosis method" to determine whether the residual coefficients need to be divided.In order to obtain the optimal residual coefficient judgment condition for any probability CU,we use the least square method to study the relationship between the probability value of the coding unit and the residual coefficient distribution.The experimental results show that,when the coding efficiency loss is small,the coding time is reduced by 59.33% on average.Secondly,this paper proposes an intra-frame CU partition prediction algorithm based on dual-path convolutional neural network.Because the "skewness kurtosis method" is difficult to accurately determine the distribution of the 8 8 residual coefficients,this paper proposes an intra-frame CU partition prediction algorithm based on the dual-path convolutional neural network classification model.We first obtain the probability of 8 8 size CU through the dual-path convolutional neural network classification model,and then calculate the judgment condition of some all-zero blocks through statistical distribution,and study the judgment conditions of the corresponding partial all-zero blocks according to the probability of the coding unit Finally,the "least squares method" is used to establish the functional relationship between the probability value and the statistical parameter.Experiments show that the coding time is reduced by 27.40% on average when the quality and efficiency loss is small.Finally,the two algorithms are integrated,and the comprehensive experiment shows that the total coding time is reduced by 63.00% on the premise of ensuring that the quality loss is within an acceptable range.