2

So, I have a cGAN model that generates random images like the ones below. Using python's ssim metric, I'm trying to generate an image that is at least 80% similar to the one on the left. I have generated bout 20,000 images and the maximum similarity I've seen is 76%. The images below received 76% similarity.

Is there a similarity metric that would be better suited for this purpose: I'd like to compare the black structure in the image and not necessarily its location in space. For instance, if I generated an image that looked exactly like the left image (same size and location of holes, same outline of the structure) but it was located higher or lower in the 64x128 pixel space, I would want that to be at least 90% similarity. Is there a metric like this?

Thanks.

enter image description here

srv_77
  • 547
  • 1
  • 8
  • 20
  • Image similarity is not a single universal concept. Specifically SSIM addresses just one kind of similarity. Here is a page to help understand similarity better: https://vitali-fedulov.github.io/similar.pictures/research.html. Deep learning approach helps with "not well defined" similarity by training a neural network for your specific sets. You can experiment with a specific image set in Google Teachable Machine: https://teachablemachine.withgoogle.com/ – Similar pictures Jul 08 '22 at 20:27

0 Answers0