Originally Posted by rowan

I've been working on and off on something like this, so far it can recognise images that have been modified slightly (eg resizing, resaving, minor changes in colour and saturation) but it cannot handle crops.

What is really needed is detailed image analysis to create several keys, with each describing small parts of the image, eg a 3 pixel triangle, or two blobs that are X distance apart. Keys from the copyrighted image are compared with keys in the database, and the best matches throw a result.

This is not simple by any means, and it will also detect images that are very similar (for example two cropped images of the same model with a similar facial expression in each)