Research And Application Of No-reference Video QoE Evaluation Method For Streaming Media

The rapid development of mobile communication and streaming media technology in recent years has promoted the explosive growth of various video services,making it more and more important in people ' s lives and work.Quality of Experience(QoE)is an important indicator to reflect users ' satisfaction with the service and an important basis to judge the quality of streaming media information dissemination services.Although deep learning technology,which has been widely used recently,has promoted the progress of this field,existing research work still faces many problems when applied to practical scenarios.First,the existing video quality databases are all oriented to video on demand(Vo D).Today,with the gradual popularization of real-time audio and video applications(such as video conferencing),the relevant QoE subjective evaluation data set is still very scarce.Secondly,the existing non-reference QoE model with high complexity is difficult to meet the needs of QoE real-time monitoring;in addition,in the face of complex and changeable streaming media content in application scenarios,the generalization of existing methods is poor.To solve the above problems,this paper builds a QoE subjective evaluation data set for video conference scenes;then a low complexity QoE prediction model is proposed from the perspective of fluency modeling.Finally,a high generalization QoE prediction model is proposed from the perspective of feature decoupling.The main work and innovation of this paper are as follows:(1)In terms of data sets,a large-scale subjective annotation dataset for video conference scenes is built.recording source video in two common usage scenarios of video conference;then the network loss instrument is used to simulate various network conditions to generate distorted video;then subjects ' subjective ratings of distorted video are collected through crowdsourcing scoring platform.Finally,the collected scores are cleaned to form a subjective annotation dataset for video conference scenes,and a brief statistical analysis is carried out.(2)In terms of improving the computational efficiency of the model,from the perspective of the principle of video fluency perception,an efficient QoE evaluation method for non-reference streaming media video based on nearest neighbor frame relationship modeling is proposed.Through pixel level and frame level threshold learning parameters,a feature extraction operator with strong correlation of perceptual damage is designed to replace the loop neural network module to reduce the computational redundancy of the model.At the same time,the end-to-end training of spatial and temporal dual branch modeling is realized to improve the computational efficiency and prediction accuracy of the model for video damage perception reasoning.(3)In terms of improving the generalization of the model,from the perspective of eliminating the influence of video content changes,a robust QoE evaluation method for non-reference streaming video based on depth visual feature decoupling is proposed.Through the residual module and correlation calculation module,the typical correlation analysis and confrontation learning strategy are applied to separate the attributes of the depth features extracted by the backbone network in the cyclic confrontation training stage,so as to reduce the perceptual quality-independent features extracted by the model,and improve the generalization and prediction accuracy of the model for video damage perception reasoning in the changeable application scenarios.