mage-based search specialists TinEye might not even have left closed beta yet, but it looks like they've got a fight on their hands in the new search battleground thanks to a team of researchers at MIT.
The group, based at the Computer Science and Artificial Intelligence Laboratory and lead by Antonia Torralba, has been working on reducing the amount of information required about an image in order to still be able to make judgements regarding its content. So far, their work has produced some interesting results.
In order to perform image recognition you first need to boil a picture down into a mathematical 'hash' that contains enough information to compare with other images. The bigger the hash, the more likely you are to find what you're looking for ? but the larger the database required, and the slower the search. Accordingly, Torralba's team is ?trying to find very short codes for images,? in order to boost the performance of image-based searching.
The team's work has discovered that a hash as small as 256 to 1024 bits in length is enough to track images across a large corpus. By creating the tiny hashes ? which can be calculated without human intervention and work on the entire image rather than individual sub-sections ? Torralba's team was able to reduce a set of 12.9 million images gathered from the Internet to a database just 600MB in size ? small enough that an average desktop PC is capable of searching the entire collection for a match in seconds.
Torralba coughs to problems recognising complex or unusual shapes via the system, stating that ?not all images are created equal.? Common objects ? recognising the shape of a person, or a flower or building, for example ? seem to be picked up quite easily given a large enough sample set. Torralba believes the research demonstrates that ?with very large amounts of images, even relatively simple algorithms are able to perform fairly well.?
Facebook
Twitter