Geoffrey Hinton, University of Toronto
Image Retrieval using Short Binary Codes
The obvious way to find images that are semantically similar to a
query image is to solve the object recognition problem. In the
meantime, it is possible to extract a feature vector from each image
and to retrieve images with similar features. If the features are
binary they are cheap to store and match. If they are also highly
abstract (e.g. indoor vs outdoor) and roughly orthogonal they will
work well for image retrieval. I will describe a method of extracting
such binary features using deep belief networks. I will then
show how binary codes can be used retrieve a shortlist of semantically
similar images extremely rapidly in a time that is independent of the
size of the database.
This is work in progress with Alex Krizhevsky.