Improving the prediction accuracy of PSNR by simple temporal pooling

Abstract

PSNR is still one of the most often and universally used visual quality metrics. Although it is not very well suited to describe the human perception of visual quality, its simplicity and familiarity lead to its extensive use in many applications. We propose to improve the predication accuracy of PSNR by simple temporal pooling and thus not only using the mean PSNR, but also to exploit other statistical properties. In order to support this approach, we conducted extensive subjective testing of HDTV video sequences at typical bit rates for consumer and broadcasting applications. Using temporal pooling, we were able to achieve an improvement of nearly 10 % in the predication accuracy of PSNR for visual quality while not increasing the computational complexity significantly. Also this approach may be extendible to other frame-based metrics.