| Summary: | Properties of a popular measure of comparing a digital image with a reference – the index of structural similarity, called SSIM in the literature – are explored. It is proved that the SSIM and its derivative functions are not metrics. Variants of the index modification are described. It is shown that measures similar to this index evaluate not quality of images, but their similarity by fragments. Additionally, it is shown that the averaged expert assessments called MOS are very subjective and cannot exact correlate with numerical estimates of similarity of the compared images. To get the SSIM index, a matrix of local estimates is calculated. Each evaluation determines similarity of two analyzed pixels with the same coordinates taking into account neighboring pixels. Usually, the average arithmetic value of the obtained matrix is used as the SSIM index. Instead, to improve the SSIM index, it is proposed to use the scale parameter of the Weibull distribution, which approximates the histogram of the local index estimates. On a set of images from the public database TID2013, it is shown that the proposed parameter is a more accurate characteristic of image similarity than the mean of local estimates.
|
|---|
