VR Video Quality Assessment Method Based On 3D Convolutional Neural Network

Virtual Reality(VR)is an interactive computer-generated experience taking place within a simulated environment.In recent years,VR technology has developed rapidly,and the number of VR videos has multiplied.It is of great significance to study the method of VR video quality assessment(VQA).VR VQA is divided into subjective assessment and objective assessment.The factors affecting users'perception of VR video quality are not only the perceived quality of video,but also the subjective feelings such as presence,cybersickness and acceptability.However,there are few subjective quality assessment database of VR video,among which there are even fewer studies on the sense of presence,cybersickness and acceptability.Therefore,this paper establishes a subjective quality assessment database of VR video,which not only includes the score of perceived quality,but also includes the score of presence,cybersickness and acceptability.Subjective assessment is time-consuming and there is no objective assessment standard for Panoramic video so far.In this paper,the method of VR VQA is analyzed in detail,and a new non-reference assessment method of VR video quality is proposed.The method uses 3D convolution neural network(CNN)to train VR video in two steps,namely,classification training and regression training,so as to calculate the quality score of VR video.Because of the time-consuming two-step training,this paper improves the network structure,and directly uses regression training to evaluate the quality of VR video,and finally finds that the effect is good.In addition,this paper uses two cutting methods,namely the ordinary cutting method and the viewport cutting method,to expand the VR video library and use it as the input of the VR VQA network,so as to compare the influence of the two methods on the training results,and also compare the impact of the input size of the neural network on the training results.The prediction results obtained by this method are highly consistent with the subj ective quality score of VR video,and the prediction effect is better.Compared with the existing VR VQA methods,it has strong competitiveness.In the future,we could test this VR VQA method on more databases and improve it.