A new psychovisual paradigm for image quality assessment: from differentiating distortion types to discriminating quality conditions

This paper investigates the impacts of image quality level on the prediction accuracy of image quality metrics. While many state-of-the-art perceptual image quality assessment methods have achieved fairly well performances in terms of the correlation between the quality predictions and the subjective scores, none of them took into account the effects of the quality levels of those test images on prediction accuracy of the quality metrics. In this work, inspired by the mechanism of human perception under high- and low-quality conditions, we propose a new image quality assessment paradigm based on image quality level classification. Our investigation on TID2008 and other three publicly available databases (LIVE, CSIQ and Toyama-MICT) results in two valuable findings. First, the performances of major well-known image quality assessment methods are significantly affected by image quality level. Second, through combining different quality metrics for different quality levels, superior performance can be achieved as compared to some of the best image quality metrics, e.g., SSIM, MS-SSIM, VIF and VIFP. Experiments and comparative studies are provided to confirm the effectiveness of the proposed new paradigm by differentiating quality levels for image quality assessment.