COME for No-Reference Video Quality Assessment

Abstract

Nowadays, the issue of objective Video Quality Assessment (VQA) has been extensively studied. In this paper, we present an effective general-purpose VQA method named COnvolutional neural network and Multi-regression based Evaluation (COME). It requires no referred lossless video and is universal for non-specific types of distortion. A modified 2D convolutional neural network is introduced to learn the spatial features at frame level. At the same time, the motion information is extracted as temporal features at sequence level. And a multi-regression model is proposed to comprehensively assess the final video quality according to human’s psychological perception. The proposed method is tested on two commonly used databases with numerous kinds of distortions. The experimental results show that the proposed COME method is comparable with most popular full-reference VQA methods.