Research On Video Quality Enhancement Algorithm Based On Deep Learning

With the rapid development of the Internet,video has become an important part of our daily work with its most intuitive feature of recording things.The defect of its large storage volume has brought tremendous pressure to network transmission and storage.The current mainstream video coding compression standard HEVC has been widely used and can achieve good video compression efficiency,However,in the face of the increasing total amount of digital video and the rich application scenarios in real life,the low bit rate and high compression performance pursues will cause serious compression distortion of the video,such as ringing effect,block effect.Therefore,how to improve the subjective quality of the encoded video under the condition of low bit rate or to improve the subjective quality of the low-resolution video is an important issue that needs to be solved urgently.At the same time,deep learning has developed rapidly in the field of video image processing,and has achieved many achievements that traditional algorithms cannot achieve.How to effectively apply deep learning technology to the field of video image processing to solve practical problems needs further exploration.Based on the above research background,this article proposes a new deep learning-based video compression distortion repair algorithm to improve based the video quality after encoding at low bit rate and a new deep learning-based video super resolution algorithm to improve the video quality at low resolution.The main research results are as follows:First,in view of the problem of video quality at low bit rates,a convolutional neural network algorithm that effectively combines the video frame time domain information with a video encoder is proposed.The algorithm first uses the multi-scale attention module to extract the multi-scale detailed feature maps in the video frame,and then extracts the block division information of the coding unit and transform prediction unit generated during HEVC low-bit-rate encoding as the distortion repair guide map,and finally the detailed feature map and distortion repair guide map are input into the Conv LSTM network in chronological order,and Conv LSTM simultaneously performs sequential and reverse order bidirectional iterations to effectively fuse time-domain inter-frame information and complete the final repair of video frames.Our proposed algorithm can improve the average PSNR value of distorted video by 0.39 d B.At the same time,our algorithm also effectively reduces various compression distortions in video in subjective tests.Secondly,in order to solve the problem of low-resolution video quality,a convolutional neural network algorithm combining inter-frame motion compensation with time-domain information of video frames is proposed.The algorithm first performs displacement calibration on adjacent frames of the input LR through a motion compensation network,and then uses a two-way Conv LSTM network to extract and fuse features of the calibrated adjacent frames and LR,and finally performs sub-pixel convolution on the output feature map to get HR.Compared with other super-resolution algorithms,our proposed algorithm has higher subjective and objective quality under various magnification tasks.Compared with the most commonly used Bicubic interpolation algorithm,our algorithm can brings average 4.89 d B benefit of PSNR in 2 × task and also brings richer and more accurate detailed information to the sub-video.