Representation of images for classification with independent features

Abstract

In this study, independent component analysis (ICA) is used to compute features extracted from natural images. The use of ICA is justified in the context of classification of natural images for two reasons. On the one hand the model of image suggests that the underlying statistical principles may be the same as those that determine the structure of the visual cortex. As a consequence, the filters that ICA produces are adapted to the statistics of natural images. On the other hand, we adopt a non-parametric approach that require density estimation in many dimensions, and independence between features appears as a solution to overthrow the curse of dimensionality. Hence we introduce several signatures of natural images that use these feature, and we define some similarity measures that correspond to these signatures. These signatures appear as more and more accurate estimations of densities, and the associated distances as estimations of the Kullback-Leibler divergence between the densities. Efficiency of the couple signature/distance is estimated by a K-nearest-neighbour classifier, with a leave-one-out procedure for all the signatures we define, and a bootstrap based one for the best results.