Research On Neural Network-Based Chroma Coding Methods

In the era of information,the storage cost and bandwidth requirements of video data continue to skyrocket.Therefore,video coding technology has important eco-nomic value to the industry,and more efficient video data compression algorithms have become the top priority of current video technology development.In 2020,the Inter-national Standardization Organization jointly released the latest video coding standard H.266/VVC(Versatile Video Coding)standard to improve the coding efficiency of ultra-high-definition video.The VVC video coding standard adopts a hybrid coding frame-work,which mainly includes several important modules such as prediction,transform,quantization,entropy coding,and loop filtering.In the development of video coding standards,a variety of intra or inter prediction methods are used to improve the coding performance of the luminance component,and compared with the luma component,the chroma component has received relatively less attention.For this reason,this thesis mainly conducts in-depth research on chroma coding in video coding,discusses the deficiencies of VVC coding methods,and proposes more efficient and novel methods.This thesis proposes neural network-based chroma coding,including prediction and interpolation.The main work is:1.Neural network-based cross-component chroma prediction.In HEVC standard,the cross-component linear model(CCLM)is proposed as a linear prediction model to achieve efficient chroma prediction.CCLM establishes a linear relationship between the reference sample of the chroma block and the reference line sample of the co-located luma block,which is used to realize the linear prediction from the luma to the chroma component.According to the principle of CCLM,we propose a cross-component pre-diction neural network(CCPNN).Experiment results show that this method has better coding performance than the CCLM method.Under the All-Intra coding configuration,the three components of YCbCr are improved by an average of about 0.45%,1.72%and 1.28%respectively.2.Neural network-based chroma inter-polation.In the video coding process,it is inevitable to down-sampling the chroma components.Aiming at the chroma compo-nents after the down-sampling,this paper refers to the principle of fractional pixel in-terpolation in inter prediction,combined with the cross-component chroma prediction method,and proposes a framework for chroma interpolation based on neural network.The framework includes invertible fractional pixel interpolation based on neural net-work,and the cross-component chroma prediction network proposed in the first work of this thesis.Through interpolation and prediction of the compressed image,its objective quality is improved to a certain extent,which verifies the feasibility and effectiveness of this scheme.