Research On No-reference Natural Distortion Video Quality Assessment Method Based On Deep Learnin

Nowadays,with the rapid development of Internet technology,video is used as a carrier in all walks of life,providing convenience for people’s life,work and study,and the sharing and transmission of video has become more frequent,so the quality of video plays an important role in people’s acceptance of information.In the process of capturing,compressing and uploading videos,the video quality is degraded due to distortion.Therefore,it is necessary to evaluate the video quality in order to ensure that people can get high quality viewing experience and video service at the terminal.For the natural distortion that occurs in operations such as capture,compression and upload,i.e.,user-generated content videos.In this paper,we combine visual perceptual properties and use deep learning methods to investigate video quality assessment.The main research components and overview are as follows:(1)A no-reference video quality assessment based on deep network and visual perception is presented.Combining attention mechanism and visual perceptual features for quality evaluation,firstly,a deep network combining convolutional neural network and attention mechanism is constructed to extract depth perceptual features for framelevel images,and covariance pooling is incorporated in the down-sampling of features to extract second-order information of features;secondly,a Transformer network is used for temporal modeling to learn the long-term dependence of perceptual quality prediction.Finally,a visual perceptual weighting strategy is used to weight and sum the frame-level scores to obtain the final video quality scores.The experimental results demonstrate that the method can effectively model the perceptual process of video quality,and the objective prediction values have high consistency with the subjective perception values.(2)A no-reference video quality assessment based on spatio-temporal perception feature fusion is presented.This method is an improvement to the problems of method(1),constructing a dual-branch convolutional neural network from distorted video frames and frame difference maps generated from a global perspective,incorporating a motion-guided attention module,considering the interaction between spatial and temporal information,and fusing spatio-temporal perceptual features from a multi-scale perspective;furthermore,introducing a temporal modeling network for further longterm sequence fusion,and finally using average temporal fusion to obtain the final perceptual quality score.The objective prediction value has a high consistency with the subjective perception value.(3)A no-reference video quality assessment for natural distortion based on dualbranch 3D convolutional neural network is presented.Considering the spatio-temporal characteristics of the video,a simple 3D convolutional neural network is constructed to extract the spatio-temporal perceptual features in the video,and the gradient stream is used as an auxiliary to extract high-frequency information such as edge and structure,reducing the difficulty of extracting features from a single stream.Then,the model is trained using an auxiliary loss function that minimizes the sum of predicted and true values to improve the model performance.Experiments show that the method has good agreement with subjective perception.