Research And Application Of Video Quality Assessment Method Based On Deep Learnin

With the rapid development of mobile devices and social media platforms,video has gradually become an important channel for people to learn,work,and play.Before people can view video,they need to undergo processes such as production,encoding,compression,network transmission,reconstruction,and display,which often lead to distortion and degradation of the video,affecting people’s viewing experience.In order to ensure the visual experience of end users,research on video quality assessment is very important.Effective video quality assessment methods are urgently needed throughout the entire process from video capture,compression,to video transmission,and reconstruction.In response to the above requirements,based on whether there are original videos in the assessment process,the paper studies two types of video quality assessment methods,full reference and non reference,respectively.The specific content is summarized as follows:(1)Considering that video distortion mainly comes from degradation of video quality caused by spatial and temporal distortion,a full reference video quality assessment method STPFVQA combining spatiotemporal characteristics and visual perception is proposed.Firstly,a convolutional network is used to extract spatial perceptual features from reference and distorted videos;Secondly,the extracted spatial awareness features are fed into the Transformer codec to construct the serialization relationship of the video,while comparing the reference video and the distorted video to explore the impact of distortion on the video sequence relationship;Then,the output of the Transformer is fed into the prediction header to form a frame level score;Finally,in order to simulate the perception lag of the human visual system,the final video quality score is obtained by comprehensively considering short-term,long-term,and global memory effects.The experimental results show that the proposed model conforms to the perceptual situation of the human visual system.(2)Considering that there is no original video as a reference for reference free video quality assessment,a reference free video quality assessment method JTA-NRVQA combining JND graph and dual attention mechanism is proposed.Firstly,the JND map of each frame of distorted video is extracted as auxiliary reference information,and the spatial features of distorted video and JND map are extracted using the EFFICIENT network.The extracted spatial features are globally averaged and sent to a fully connected structure.The features obtained from JND map are used as perceptual weights for non reference video frames,and the two are multiplied to obtain a frame level spatial quality score;Secondly,Vit network is used to fuse the spatial features of distorted video and JND images to make the extracted features conform to the characteristics of the human visual system.Transformer encoder structure is used to serialize and extract the features to form a frame level temporal distortion degradation score;Then,the spatial and temporal degradation scores are added and averaged to obtain a frame level spatiotemporal degradation quality score.Finally,the frame level spatiotemporal degradation quality score is averaged to obtain a video quality score.Achieved relatively advanced performance on three non reference video quality assessment datasets,Ko NVi D-1k,CVD2014,and LIVE-VQC.(3)Design and implementation of video quality assessment method system.Based on the above algorithm research,this paper designs and implements a visual video quality assessment system.This system builds a video platform that integrates video quality assessment,video playback collection,and other functions to provide users with online video quality assessment functions.