| With the increasing demand for full HD and ultra HD video,video codec technology is also in urgent need of improvement.Although High Efficiency Video Coding(HEVC)greatly improves the coding performance,at the expense of high computational complexity,which is difficult to apply and deploy in real-time scenarios.Therefore,how to reduce the complexity of HEVC has become an urgent problem to be solved,which is the primary content of research in this paper.This paper simplifies and accelerates the Coding Tree Unit(CTU)partition process based on the coding characteristics,a dataset containing CU neighborhood information and image information is created,and proposes two fast coding algorithms based on the dataset to adapt to different application scenarios.The primary research methods and innovative results of this paper are as follows:1)A lightweight network-based fast Coding Unit(CU)partitioning algorithm L-CNNs is proposed.The algorithm converts the Rate-distortion Optimization brute force search process of the original CTU partition into a three-level cascaded binary classification problem,and uses L-CNNs combined with neighborhood information to predict the CU partition mode layer by layer to obtain the final CTU structure.L-CNNs effectively save coding complexity with little coding performance degradation.(2)A fast CU partitioning algorithm GL-CNNs based on grouped convolution is proposed.The algorithm uses grouped convolution to design the network structure,and combines the channel rearrangement module to enhance the feature interaction between different groups,so that GL-CNNs can achieve almost the same encoding performance as L-CNNs and further reduce the amount of parameters.In addition,by improving the structure of the second-layer classifier,the GL-CNNsV2 algorithm is proposed to make the network more lightweight.Experiments show that the proposed L-CNNs algorithm based on lightweight network can save 60.46%coding time on average and increase BDBR by only 2.78%,with a very small amount of model parameters.The GL-CNNs algorithm based on grouped convolutions can achieve almost the same coding performance as L-CNNS when the number of parameters is reduced by 45.88%.It is further proposed that GL-CNNsV2 can reduce the number of parameters by 64.52%and make the network more lightweight,and the overhead of introducing Convolutional Neural Network is extremely small,making HE VC suitable for many different application scenarios. |
