Research On Universal No Reference Video Quality Assessment Method

With the rapid development of multimedia and Internet technology,videos are widely spread in various applications including digital TV,video website,video conference and so on.However,due to the limit of video compression technology and transmission channel,various distortions are unavoidably induced into videos,which can significantly influence users' quality of experience.Therefore,in order to improve the quality of service,it is very necessary to assess video quality accurately.Objective quality assessment methods aim to build mathematical assessment models by which machine can score the quality of videos automatically without involvement of human raters.Among them,universal no reference video quality assessment deserves further study because it doesn't need reference videos and applies to different distortion types.This paper focuses on universal no reference video quality assessment methods,and the main work is divided into three parts.Firstly,the current research status of video quality assessment at home and abroad is introduced in detail,and the principle and performance of some classical no reference algorithms are also analyzed and compared.Secondly,on the basis of foreign advanced algorithm,this paper proposes an improved algorithm based on two-stage temporal pooling model,which is more accurate and stable than the original one.Last but not least,on the basis of intensive study of deep learning,this paper applies deep learning model to video quality assessment and proposes an algorithm based on 3D convolutional neural networks,and the proposed algorithm has low computational complexity and high assessing speed with high accuracy,which can maintain stable performance when the video content and distortion type change and thus is very practical for real applications.