Using STRESS to compute the agrement between computed image quality measures and observers scores: advantages and open issues

Abstract

[EN] Visual quality of color images is studied through costly psychophysical experiments, which are used to record observers quality scores. Visual image quality metrics pursue to maximize the agreement between computed quality and observers scores. Therefore, it is of critical importance to have appropriate measures for this agreement, both for the development and use of the image quality metrics. The most used one is the well known Pearson correlation coefficient while Spearman rank correlation coefficient is also customary used. In this work we explore the use of an alternative metric: The standardized residual sum of squares (STRESS). STRESS has some interesting properties that encourage us to use it for measuring the agreement between computed image quality and observers scores, being the most important one the possibility to run statistical significance tests between metrics. We will compare the performance of STRESS with Pearson and Spearman coefficients using both synthetic datasets as well as a recent visual image quality evaluation dataset. As it will be shown, the performance is different and we found several points in favor of using STRESS along with some interesting open issues.