I have around 3 million BW images. I would like to organize them in as few clusters as possible in a way that is meaningful for the dataset without any prior knowledge for this data, as they come from domains that are not represented in classic models trained in classic datasets (like ImageNET).
My target is to produce meaningful labels for images. By meaningful I mean in a machinic way. I don't want exactly to have human-readable labels, but I would like for the labels not to be that arbitrary (thus to relate with higher-semantically or deep correspondences from the vast amount of pictures, which I supply to the model).
Are you aware of any method able to do something like this?