Pretty interesting. I've played with this idea as well, to make a lexicon of image tiles, and then puzzle them together to achieve compression/recognition of l

Pretty interesting. I've played with this idea as well, to make a lexicon of image tiles, and then puzzle them together to achieve compression/recognition of large sets of images and video data. I wonder, though, how well does this work in practice, for photos and illustrations? Deduplication works great for the specific application of screenshots, but there is almost a zero chance that the tiles from two different photos match exactly. You'd need some kind of fuzzy matching, which will greatly increase the computational cost of the algorithm.