Research And Implementation Of No Reference Stereoscopic Video Quality Evaluation Method Based On Deep Learning

With the continuous development of stereo technology,stereoscopic video-related applications and service systems are beginning to approach people's lives,playing an increasingly important position.In these stereoscopic video systems,controlling the perceived quality of video content is the key to enhance the user experience.However,in various processing stages of the 3D video system,3D video may suffer different degrees of damage due to limitations of bandwidth,equipment,and technology,so that the performance of the 3D video system is greatly affected.For the large amount of stereoscopic video data in the system,it is necessary to apply the stereoscopic video quality assessment(SVQA)to evaluate and quantify the visual quality when monitoring,controlling and improving the stereoscopic video quality.Most of the traditional methods for evaluating the quality of unreferenced stereo video have similar frameworks.Divide-and-conquer method is usually used to simplify the problem,including manual feature extraction and classifier design.But in reality,we must consider the non-intuitive interaction of complex visual factors such as video content information,depth perception information and time-domain information when evaluating video quality.Therefore,it is difficult to design a set of features suitable for various types and degrees of distortion.In order to solve the shortcomings of the traditional methods,we have introduced the convolutional neural network into the field of stereoscopic video quality assessment,and established a no reference stereoscopic video quality evaluation framework based on 3D convolutional neural network.Firstly,we complete the preprocessing and data enhancement of stereo video based on human stereoscopic characteristics and quality evaluation related features.Then we design a 3D convolutional neural network structure,which uses local short video cube as input to extract local spatiotemporal information simultaneously.And then quality score fusion strategy that considers global time clues is used to obtain the final video quality prediction score.In addition,we use the 3D convolutional neural network as the feature extractor to further improve the linear correlation between the predicted quality score and the subjective evaluation score by replacing the fully connected layer in the original network with support vector regression to complete the mapping from feature to quality score.Experiments show that the effect of this method is significantly better than that of the most advanced method at present.In addition,the proposed method does not depend on complex pre-processing,and the network structure is simple.When using GPU acceleration,the computational efficiency is more advantageous and practical.