Research On Video Coding Structure And Optimization Algorithm

With the development of Internet information industry and cultural entertainment industry,people's demand for the clarity of video data is increasing day by day.Under the premise of limited bandwidth,keeping high resolution and high quality of video data is the main research target of video coding standard.Inter-frame prediction reference and rate-distortion optimization in video coding are two extremely important technologies.Through the inheritance of the correlation between video frames and the combination of the rate Lagrangian optimization theory,the quality and rate of video coding are well balanced.The existing reference structure and optimization parameters adopt empirical training parameters,which are not adaptive to diversified video information sources.The fixed parameter assignment does not fully consider the time-domain dependence between reference frames.The empirical optimization parameters are not complete in theory and cannot be further approximated to the optimal value.Combined with convolutional neural network method,the problems of video coding reference structure and optimization method are studied from three aspects.1.An adaptive reference frame selection algorithm is proposed to solve the problem of low delay coding structure's lack of adaptability.According to the low-delay reference structure,the dependence of the coding bit is counted,the influence factor model of the reference frame is established,the maximum radiation distance of the reference frame is calculated,and a reasonable reference set is allocated for the adaptive encoding frame to improve the encoding performance.Compared with HEVC reference software,the proposed algorithm achieve at most 0.93% compression performance improvement.2.Aiming at the problem of fixed parameters of video time-domain hierarchical structure,a reference structure-determined Lagrange multiplier sub-algorithm is proposed.According to the characteristics of time domain hierarchical structure,a hierarchical reference strength model was constructed.In combination with the differences in time domain hierarchy,reference strength and quality,a hierarchical Lagrange multiplier adjustment model was established to allocate reasonable optimization parameters for time domain stratification and obtain coding performance gains.Compared with AVS2 standard reference software,maximum 1.3% coding performance improvement is achieved under various coding structures.3.According to the convolutional neural network method and the receptive field theory,a receptive field decline convolutional neural network is designed to describe the image features in depth,coarse-grained and fine-grained.Iterative training and optimization of network structure can improve the quality of distorted images.The network was applied to the ultra-high resolution image and achieved the comprehensive PSNR improvement of 0.5db on average.Through the research mentioned above,this paper proposes adaptive reference method and encoding optimization parameter selection strategy to further improve video encoding efficiency under the current mixed coding architecture.Compared with HEVC and AVS coding standard platform,it has obvious gain in coding performance.In this paper,the receptive field decline neural network is proposed to improve the image quality of ultra-high resolution and multi-scene images.The optimization tool proposed in this paper has been adopted and applied by the AVS2 standard.The proposed network structure has participated in the deep learning image filtering competition hosted by ChinaMM and won the creativity award.