I found this a pretty interesting read - reminds me of Latent Semantic Indexing, extracting the general concepts from the images and using them to classify new unseen data. LSI extracts concepts from textual information and can then classify new documents with a fairly high accuracy. Basically low value words are discarded, whilst high value, relevant keywords are kept and ranked.

http://www.jacobsschool.ucsd.edu/news/news_releases/release.sfe?id=650