Retrieval performance improvement through low rank corrections

Whenever a feature extracted from an image has a unimodal distribution, information about its covariance matrix can be exploited for content based retrieval using as dissimilarity measure, the Bhattacharyya distance. To reduce the amount of computations and the size of logical database entry, we approximate the Bhattacharyya distance, taking into account that most of the energy in the feature space is often restricted to a low dimensional subspace. The theory was tested for a database of 1188 textures derived from VisTex with the local texture being represented by a 15 dimensional MRSAR feature vector. The retrieval performance improved significantly, relative to the traditional Mahalanobis distance based approach, in spite of using only one or two dimensions in the approximation.