Sparse-distributed codes for image categorization

This paper addresses the problem of image categorization using local sensory information which is aggregated into global cortical-like representations of different image categories. Local information is adaptively extracted from an image database using independent component analysis (ICA) which provides a set of localized, oriented, and band-pass filters selective to the most independent features of the different categories. Such local representations have been computationally investigated by several researchers, and have also been experimentally observed as characteristics of simple cell receptive fields in the primary visual cortex. However, little work has been done on further use of these representations to provide more complex and global description of images. In this paper, we present an algorithm which uses the energy of a minimal set of filters to provide category-specific signatures which are shown to be strongly discriminant. Computer simulations are carried on an image database consisting of three categories (faces, leaves, and buildings). The categorization performances of the algorithm using ICA and PCA filters are reported. It is mainly shown that considering a small number of PCA filters leads to a performance which is not significantly improved by considering other PCA filters, however, considering additional ICA filters increases performance due to the fact that each additional filter carries additional information (in the entropy sense).