Research On Deep Learning-based Video Intra Prediction And In-loop Filtering Method

The uncompressed raw video signal contains a huge amount of data,and it is impractical to directly store and transmit the raw data.Advances in video coding have enabled us to deliver higher-quality videos through bandwidth-limited channels.With the rise of the Internet and the popularity of ultra-high-definition devices,the number and resolution of video are exploding.Although modern video coding technology can effectively compress video,it still cannot satisfy the demand for more higher-quality videos.Recently,with the availability of powerful GPUs and the large amount of Internet data,deep learning has shown great potential and has achieved great success in many fields.The application of deep learning to video coding to further improve compression efficiency has attracted widespread attention and become the frontier of academic research.This thesis mainly studies deep learning based video coding technology,the main work can be divided into four parts: First,the current situation and development trend of deep-learning based video coding are comprehensively introduced according to the modules in the video coding framework.Especially,the excellent algorithms of deeplearning based intra prediction and in-loop filtering are introduced in detail.Secondly,a highly versatile recursive residual convolution network(RRCNN)is presented.The network combines recursive learning strategies with residual learning strategies and can be widely used in a variety of image enhancement and restoration tasks..Thirdly,in view of the fact that the traditional intra prediction cannot make full use of context information and the linear predictor for objects with directional structures is too simple,an intra prediction method based on recursive residual convolutional network is proposed.The network takes both the adjacent reconstructed block and the traditional prediction block as the input,and predicts the residual between the traditional predicted block and the original block.The specific structure of the network consists of three parts,the trained single multiplexing network model can process the different-sizes blocks,reducing the quantity of model parameter.The proposed method achieves 2.8%bitrate saving.Fourthly,recursive residual convolution neural network based in-loop filtering for intra frames is proposed.The network adds quantization parameter information to the input of the network to enhance the generalization ability.The trained single model can widely process different-quality reconstructed frames and greatly improves the quality of filtered videos,achieving 8.7% bitrate saving.