Here are two aligned images (greyscale, but same applies to RGB) that I want to compare using SSIM:

real fake

Clearly, these are very similar images. It seems that the first image is more blur rr However, the SSIM is only 0.52. I would expect it to be higher. Is SSIM not a good metric for measuring similarity in this case? Why is this the case?

  • 1
    $\begingroup$ "clearly, these are very similar images", um both I and your arbitrarily picked measure for similarity would like to heartily contradict you. $\endgroup$ Commented Mar 23, 2020 at 10:43

1 Answer 1


clearly, these are very similar images

Both I and your arbitrarily picked measure for similarity would like to heartily contradict you, there.

It was you who picked SSIM, Structural SIMilarity, as measure; if that measure doesn't describe your own idea of similarity well enough, well, you might want to define what similarity is and come up with a different measure.

Here, specifically: SSIM is especially designed to respect how focused, sharp, high-contrast images are. That's the whole point of it: although a low-pass version of your two images would indeed be very similar, the SSIM, meant as perceptive quality index, should identify that the second one is very different.

So, you picked pretty much the worst metric for your idea of similarity, and now you're "complaining" that it does what it's designed to do!

  • $\begingroup$ Thank you for your answer. In my field, the latter image would be perceived to be very similar to the former, and would be very close to being correct. I am trying to convert (RGB) images of the former kind to look like (RGB) images of the latter kind and would like a metric to quantify that. Note that the colors in the former are different then in the latter. I thought that SSIM would be a good metric since it compares structural similarity of the images, rather than looking at the color. Can you please suggest any other methods of comparing these two images? $\endgroup$
    – TanMath
    Commented Mar 23, 2020 at 20:54
  • $\begingroup$ I'd really do what I've recommended in my question: low-pass filter both images, and then simply do a max cross-correlation. I really can't tell you how to put your understanding of similarity in an algorithm – I don't know it! $\endgroup$ Commented Mar 23, 2020 at 21:36
  • $\begingroup$ Ok makes sens, thanks for your feedback! $\endgroup$
    – TanMath
    Commented Mar 23, 2020 at 22:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.