Replying to this post.

The perceptual hash algo (github.com/Nr90/imgsim) wasn't perfect for my use case, but it did a pretty great job. They're a way of comparing the visual similarity of two images, where slight visual differences register as smaller numerical difference.