There's limits to how short you can make the perceptual hash. The more you compress it, the more information you lose. The ML image classification models can

There's limits to how short you can make the perceptual hash. The more you compress it, the more information you lose. The ML image classification models can be used to extract a good descriptor that can be further reduced into a compact signature. https://github.com/starkdg/pyphashml https://github.com/starkdg/pyphashml For indexing, I've had some success with distance-based indexing. Here's a comparison of some structures I used: https://github.com/starkdg/hftrie#comparison https://github.com/starkdg/hftrie#comparison Feel free to contact me, if you want to discuss this further.