handle image orientation
sometimes, images are the same but with rotation, would you consider implementing a rotation management feature parameter to handle similar images but rotated ?
For 2D affine transformation-invariant hashes (including partial image-matching) I suggest having a look at this approach: https://pippy360.github.io/transformationInvariantImageSearch/
CNNs exhibit a good degree of rotation invariance. So, you should try CNN method instead of hashes for obtaining duplicates in this case. I also encourage you to have a look at the benchmarks to guide you regarding algorithm/threshold selection for your image collection: https://idealo.github.io/imagededup/user_guide/benchmarks/