This is mostly orthogonal to the author's original point (with which I concur, having also implemented image similarity via hashing and hamming distance). There

This is mostly orthogonal to the author's original point (with which I concur, having also implemented image similarity via hashing and hamming distance). There just aren't a lot of knobs to tune using these algorithms so it's difficult if not impossible to make small changes to err on the side of reducing false positives.