I'm actually guessing not. It's probably something similar algorithm-wise to what Shazam does. Probably not a crypto-hash, but more likely a hashing of robust features (in a relative as opposed to absolute location form) detected from the image to give a "fingerprint." Such features would be robust to scaling and most compression artifact issues. EDIT: Here's a good example: http://laplacian.wordpress.com/2009/01/10/how-shazam-works/