Double-layer HDR Is Backward Compatible With Parameter Prediction In Codec Systems

Compared with traditional standard dynamic video,although high dynamic range video improves the human visual experience,the current market and most consumers still use traditional SDR displays,and HDR video cannot provide HDR vision on traditional SDR displays.effect.At the same time,HDR/SDR video services can be implemented simply by storing two versions of video files on the server side,but this requires a lot of storage resources.Another alternative is to design a backward-compatible dual-layer HDR video codec system to provide flexible playback for different devices,that is,to transmit the base layer(BL)stream to the SDR display terminal to achieve SDR effect display.The HDR display terminal transmits BL and Enhancement Layer(EL)code streams to achieve HDR effect display.The main work of this thesis is to study the estimation method of important parameters in this two-layer HDR video codec system.The main contents are as follows:(1)There are residual coding algorithms and codeword range amplification algorithms(Codeword Range Amplification,CRA)for generating EL methods in the two-layer codec system.This thesis uses experiments to show that conventional residual coding algorithms cannot improve the effect of reconstructed HDR video,while CRA algorithm makes the effect of reconstructed video significantly improved.In the CRA algorithm,an important parameter to segment the threshold point of the pixel gray value has an important impact on the quality of the final reconstructed HDR video.This thesis analyzes the shortcomings of global search method and formula derivation method to obtain threshold points,and proposes a threshold estimation method based on learning.(2)This thesis proposes a threshold estimation algorithm based on Gaussian process regression model.First,analyze more than 40,000 pieces of sample data obtained through global search,and find that the PSNR improvement effect of the reconstructed video and the appearance of the best threshold point have a strong relationship with the pixel gray value histogram,total bit rate and BL/EL ratio.Correlation,and determine the eigenvalues of the Gaussian regression model;sample resampling is used to solve the problem of sample distribution imbalance;Rational Quadratic is selected as the kernel function of the training model.Experiments show that the prediction threshold of this model is close to the optimal threshold;the prediction threshold reconstruction HDR video PSNR increases by 2.2dB,which is nearly a million times shorter than the global search method,and provides the possibility for real-time calculation of the threshold in double-layer encoding.(3)This thesis proposes a scheme based on the convolutional neural network to predict the best threshold point.Experiments show that the network prediction threshold reconstruction HDR video PSNR increased by 1.8d B,compared to the formula reasoning method video improvement effect increased by 86.5%,compared with the global search method,it also shortens the time by nearly a million times,which is also the same as the threshold value in double-layer encoding.Real-time calculations provide possibilities.