Image Categorization using Independent Component Analysis

This paper addresses the problem of image categorization using local sensory information which is aggregated into global cortical-like representations of diierent 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 lters selective to independent features of the diierent categories. Such local representations have been computationaly investigated by several researchers, and have also been experimentaly observed as characteristics of simple cells receptive elds in the primary visual cortex. However, very 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 lters to provide category-speciic 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 lters are reported. It is mainly shown that considering a small number of PCA lters leads to a performance which is not signiicantly improved by considering other PCA lters, however, considering addidional ICA lters increases performance due to the fact that each additional lter caries additional signiicant information (in the entropy sense). Several attempts have been made to set up a theory of sensory coding in the primary visual system. Most of the proposed theories, supported by biological experiments, suggest that sensory coding is an information processing strategy which aims to reduce (or transform) redundancy, but preserve information, in order to produce a more eeective internal Most of the classical coding strategies can be seen as image compression schemes. Among these is Principal Component Analysis (PCA). On the other hand, Independent Component Analysis (ICA) is a relatively recent trend in signal processing and neural networks research for blind separation of source signals