In a recent Scientific American article (“A Farewell to Keywords”, July 2006), IT columnist Gary Stix provides an update on the idea of using images to find images.
Stix highlights work underway at Microsoft and Google. Both companies appear focused on geometric methods. For example, Microsoft’s approach is based on the spatial orientation of triplets of features. Feature triplets of the imaged being scrutinized are compared with feature triplets of training images in a database. Matching feature triplets constitutes a positive search hit.
Because today’s methods are based on image metadata (data about the image such as its filename, its type, associated annotations, etc.), this image-centric approach is definitely innovative, and presents interesting possibilities for application.
However, purely geometric schemes for Googling images is the pure-text analog of Googling keyword combinations. (Googling keyword combinations with Boolean expressions.) Why? Both approaches are semantically challenged. They do not allow for context to be conveyed.
Google and others are actively working on smarter (aka semantically richer and expressive) search engines. Although purely geometric methods for Googling images comprises an important first step, smarter methods will need to have a sematically solid basis.