Research On Video Quality Enhancement Algorithm For HEVC

With the proliferation of video service types and the significant increase in consumer demand for video picture quality,video data volume has shown an explosive growth trend.With the diversification and high standardization of user demands,especially the urgent demand for ultra-high definition video resources with high frame rate,high resolution and high color reproduction,the efficient transmission and storage of video is facing great challenges.The encoding capacity of the current H.265/HEVC(High Efficiency Video Coding,HEVC)standard cannot meet the demand for high-quality compressed transmission of UHD video resources.On the one hand,because the video coding technology itself is a block-based hybrid coding method,the coding process will inevitably produce compression distortion.Although the loop filtering link of H.265/HEVC standard can alleviate its impact to a certain extent,it only filters specific artifact types in turn,and does not deal with other types of artifacts,so the recovery effect needs to be improved urgently.On the other hand,in order to cope with the huge data volume inherent in UHD video,the industry generally uses video super-resolution reconstruction technology to indirectly realize the compressed transmission of UHD video.However,the existing video superresolution reconstruction algorithm still has a large room for improvement in the overscoring effect,and the computational complexity of the algorithm is high,and the efficiency of the reconstruction needs to be improved urgently.Therefore,the research of video quality enhancement algorithms for UHD video has a long way to go.This thesis focuses on the quality enhancement algorithm at the decoding end for the realistic coding requirements of UHD video services for the international video coding standard H.265/HEVC,and seeks to effectively improve the reconstruction quality of UHD compressed video with limited bandwidth cost.The main research contents and innovative achievements of the thesis are as follows:First,a compressed video quality enhancement algorithm based on coding characteristics is proposed.The algorithm does not simply consider the compressed video quality enhancement task as a video denoising or deblurring task and aims to seek a deep coupling between video coding techniques and video quality enhancement techniques.By reflecting on the video coding process,three unique features of compressed video,namely quality fluctuation,pixel missing,and texture similarity,are identified,whereby a non-aligned quality enhancement model is designed in a targeted manner.The algorithm provides a new understanding perspective for improving the quality of compressed video using limited spatiotemporal information.Second,based on the above design ideas,a Transformer-based compressed video quality enhancement algorithm is proposed.By exploiting the modeling capability of visual Transformer for long-distance dependencies and combining the features of distinct boundaries between different kinds of pixel blocks in the compressed region and more similarity within the same kind of pixel blocks,a Transformer enhancement model that conforms to the characteristics of decoded pixel distribution is explored.The algorithm provides a new exploration direction for the combination of Transformer and compressed video enhancement tasks,and speeds up the process of lightweight compressed video enhancement.Third,in order to improve the efficiency of hyper-resolution reconstruction after decoding,this thesis inherits the idea of bidirectional propagation in the video hyperresolution domain and proposes a video hyper-resolution reconstruction algorithm based on dense fusion based on the characteristics of compressed video.By deeply mining the source information in the vertical spatial domain and the horizontal temporal domain,a fine-grained lattice-based dense fusion model is explored,which opens a new idea of video super-resolution model design.Compared with the current optimal video super-resolution reconstruction model,this algorithm achieves the same super-segmentation performance with a significantly lower number of parameters and computational effort.Finally,the quality enhancement algorithm of compressed video and the video super-resolution reconstruction algorithm are combined to explore a set of "lower bit rate,higher quality" video compression transmission frameworks.The experimental results demonstrate that this framework only requires about 1/16 of the data volume compared with the original high-resolution video,and the average PSNR(Peak Signal to Noise Ratio,PSNR)of the reconstructed video can reach 27.717 d B,which proves that the proposed framework has certain optimization effect and application value,and contributes to the large-scale promotion of H.265/HEVC standard in the practical application of UHD video.