Research On Key Technologies Of Video Coding Optimization

Rapid developments of intelligence devices and internet technologies have witnessed an exploration of various video services.Video has become an important information medium in modern society.However,the huge volume of raw video data,especially for highdefinition and ultra high-definition videos,panoramic videos and light field videos,poses great challenges for storage and transmission of video.In this case,it is urgent to study efficient video coding and optimization technologies to support relevant applications.The fundamental goal of video coding optimization is to improve the perceived visual quality of viewers under the condition of limited bit rate.In addition,the complexity limitation of actual application scenarios should not be ignored.The low complexity implementation of coding optimization is the guarantee for the practicality of relevant algorithms.This dissertation thoroughly investigates the key technologies of video coding optimization.For rate-distortion optimization of low complexity video coding,two algorithms are proposed,including a rate-distortion optimization algorithm of video coding under given computational complexity for H.265/HEVC and a global convolutional network(GCN)-based fast intra coding optimization algorithm for H.266/VVC.For video coding oriented perceptual quality enhancement,three algorithms are proposed,including a deep video compression(DVC)-based compressed video perceptual quality enhancement algorithm,a deformable convolution-and quantization parameter(QP)adaption-based compressed video perceptual quality enhancement algorithm and an attention-and QP adaptation-based compressed video perceptual quality enhancement algorithm.For light field video coding,a view angle projection-based light field video coding optimization algorithm is proposed.The main contributions of this dissertation are summarized as follows:1.Aiming at the rate-distortion optimization of low complexity video coding,the ratedistortion optimization under the given coding complexity and the rate-distortion optimization of fast coding unit partition of intra-mode are studied in this dissertation.Aiming at the rate-distortion optimization under the given coding complexity,the correlation between the coding performance and the coding complexity of H.265/HEVC is analyzed,and then the rate-distortion costs,the probabilities and complexities of candidate modes of coding tree units are accurately and quickly predicted.By calculating the gain of coding performance within the unit complexity,the frame-level complexity allocation is implemented based on the heap sorting,and the advanced rate-distortion performance under the given coding complexity is achieved.Experimental results have demonstrated that,compared with x265,the proposed algorithm can control the complexity more accurately,and the average coding performance is improved by 6.59% under the same given computational complexity.Aiming at the rate-distortion optimization of fast coding unit partition of intra-mode,a network to predict partition modes of coding units is proposed based on the GCN for H.266/VVC.After capturing the global information of coding units and using the loss function designed based on the correction of mode probability,the best partition mode of the coding unit is accurately and quickly predicted,which significantly reduces the coding complexity of intra-mode H.266/VVC while maintaining the rate-distortion performance.Experimental results have demonstrated that the proposed algorithm can save 46.94% ～ 63.08% coding time with only 0.85% ～ 1.52%performance loss on average.2.Aiming at video coding oriented perceptual quality enhancement,the perceptual quality enhancement of compressed videos based on the neural network and the QP adaptation of the perceptual quality enhancement algorithms are studied in this dissertation.Three algorithms are proposed,including a DVC-based compressed video perceptual quality enhancement algorithm,a deformable convolution-and QP adaption-based compressed video perceptual quality enhancement algorithm and an attention-and QP adaptation-based compressed video perceptual quality enhancement algorithm.By using the nearest neighbor interpolation and convolution operation to replace the deconvolution operation,the end-to-end DVC network is improved,and the checkerboard artifacts that might appear in compressed videos are completely eliminated.With the help of the generative adversarial network and under the guidance of the hybrid loss function,the perceptual quality of compressed videos is significantly enhanced.Experimental results have demonstrated that,compared with the DVC network,the proposed algorithm can enhance the perceptual quality of 12.27% on average.To adapt the perceptual quality enhancement algorithms to different coding QPs,two QP adaptive compressed video perceptual quality enhancement algorithms are proposed based on the deformable convolution and the attention mechanism,respectively.These two proposed algorithms capture the temporal correlations between adjacent video frames by the deformable convolution and attention mechanism,respectively,to improve the perceptual quality of the current video frame.In addition,both of them use QP modulation networks to achieve QP adaptive perceptual quality enhancement,saving storage and transmission costs without the performance loss.Experimental results have demonstrated that the proposed QP adaptive algorithms can improve the perceptual quality of compressed videos by an average of 47.54% and 61.15%,respectively,obviously superior to the existing related algorithms.3.Aiming at light field video coding,the structural relationship between sub-aperture images of the light field is derived mathematically,and a view angle projection-based light field video coding optimization algorithm is further proposed in this dissertation.Specifically,the disparities extracted from light field raw images are used to establish the corresponding model of light field points,and then the relative pose between light field images is estimated with the nonlinear optimization on manifold.Projecting the encoded frame to the view angle of the current frame with the estimated pose,the inter-frame coding performance of light field vides is significantly improved.Experimental results have demonstrated that the proposed algorithm can achieve accurate and robust relative pose estimation between light field images,and its application in the light field video coding optimization can improve the performance by 65.08% on average.