[ j_bs \propto \frac1B_p \left( \fracdpd\rho \right) \times \textfunction(s, \epsilon) ]
import numpy as np from skimage.metrics import structural_similarity mod-rssim
Have you implemented MOD-RSSIM in your workflow? Experiment with the root transformation and stabilization constant—you will be shocked at how much more "human" your image comparisons become. [ j_bs \propto \frac1B_p \left( \fracdpd\rho \right) \times
Standard SSIM is excellent, but it has a flaw: It is not a distance metric. It is a similarity metric. Furthermore, it is sensitive to the scale at which you analyze the image. A blurry image might have decent SSIM if you look globally, but terrible SSIM if you look at micro-textures. mod-rssim