Research On Non-reference Objective Quality Assessment Methods For H.264 Compressed Videos

There has been a rapid growth of mobile video traffic, driven by the popularity of mobile devices as well as the performance increases. If the video player can automatically evalute the quality of the video sequence, then video encoding parameters can be adapted to provide better user expirence, based on system resourse and network condition. The traditional full reference video quality assessment metrics can not be directly used since the original video sequences are unavailable in clients. So it is significant to propose a non-referenece quality assessment method to evaluate the received video stream.Based on multiple linear regression, a prediction model for PSNR is proposed by analyzing different syntax layers of bitstream as well as video textures and motion complexity. Firstly, the video encoding process is analyzed, and then the causes of the video distortion and the influence factors of PSNR can be determined. Besides, the PSNR influence factors are quantified based on the detailed analysis of the the factors' physical implication. Secondly, through analyzing the interactions between the various influencing factors, and the relationship models are proposed. Finally, two PSNR prediction models are proposed, based on the multiple linear regression analysis of various PSNR influencing factors. The first model, I-frame prediction model, considers three factors including quantization parameter(QP), bitrate and the texture. The second model, P-frame PSNR prediction model, not only considers the previously three factors above, but also the motion factor.In order to test the accuracy of the prediction models, this paper use x264 to encode different video sequences including two kinds of resolution, 1080 p and CIF. Experimental results indicate that the pearson correlation coefficient(PCC) of CIF sequences is 0.98 for I frames, and 0.96 for P frames. And the PCC of 1080 p sequences are 0.95 for I frames and 0.97 for P frames. However, for the PSNR prediction of 1080 p sequences' P frames, there are some fluctuations of the prediction accuracy between different video sequences, so the accuracy can be improved further.