Research On H.265/HEVC Coding Technology Based On Deep Learning

With the deployment and operation of high-speed information networks such as 5G cellular networks and Wi-Fi 6 local area networks,high-definition video live broadcast and short video applications have increasingly become the mainstream of network services,which put forward higher requirements for audio and video clarity and encoding and decoding efficiency.8K resolution H.265/HEVC video codec will become the mainstream.Under the same image quality,H.265/HEVC can reduce the code stream by about 50% compared with H.264/AVC.Due to the introduction of characteristic encoding technology,the encoding performance is improved,but it also brings about the problem of increased encoding complexity.The rapid development of deep learning technology and neural network technology has guided a new direction for reducing coding complexity and improving coding and decoding efficiency.Based on deep learning technology,this thesis studies the H.265/HEVC intra coding unit division technology.The main research content of the full text is as follows:Aiming at the problem of shallow neural network layers and less training data in the intraframe coding unit division algorithm based on deep learning,a Cupr Net network coding unit division algorithm is proposed.The Cupr Net network is designed based on the Res Net18 network and optimize the Res Net18 network to match the output of the fully connected layer with the minimum partition size of the encoding unit.A preprocessed data set including video brightness information and coding tree unit partition structure information is constructed.In the network training,the overall brightness information of the coding tree unit is also used to calculate the depth information of the smallest size coding unit,which avoids the traditional calculation of the rate-distortion cost function and accelerates the division of coding units.The evaluation results show that compared with the official encoder,the proposed algorithm reduces encoding time by 79.83% and has a bit rate loss of 7.943%.Aiming at the problem that the coding tree unit features is not effectively utilized,based on the classification model,a coding unit division classification network CupcNet is proposed,and a CupcNet network coding unit division algorithm is proposed.A three-layer neural network was constructed,and a preprocessing dataset containing video brightness information and encoding tree unit partitioning labels was constructed.The small convolution kernel is used in CupcNet network to enhance the nonlinear expression ability of the network and reduce the number of parameters.The evaluation results show that the proposed algorithm supports different quantization parameters,and compared with the official encoder,the encoding time is reduced by 64.01%,with a bit rate loss of 2.9493%.