Structural similarity image quality reliability: Determining parameters and window size

The need to obtain objective values of the quality of distorted images with respect to the original is fundamental in multimedia and image processing applications. It is generally required that this value correlates well with the human vision system (HVS). In spite of the properties and the general use of the mean square error (MSE) measurement, this has a poor correlation with HSV, which has led to the development of methods such as structural similarity (SSIM). This metric improves the correlation with respect to the classic MSE and PSNR (peak signal to noise ratio). However, its behavior depends on the values assigned to constants and on the windows size selected. These values are usually assigned arbitrarily and there have been no studies on how they affect the SSIM. In this work, we have analyzed empirically the most appropriate values for the different constants used in the SSIM equations. We have also analyzed the importance of window size in the calculation of MSSIM, and propose a method for determining the window size based on image complexity. Using the values selected and the window size defined, the correlation between SSIM and DMOS (differential mean opinion score) is significantly improved by around 17% with respect to the values commonly used.