An Evaluation of Sparseness as a Criterion for Selecting Independent Component Filters, When Applied to Texture Retrieval

Abstract

In this paper we evaluate the utility of sparseness as a criterion for selecting a sub-set of independent component filters (ICF). Four sparseness measures were presented more than a decade ago by Le Borgne et al., but have since been ignored for ICF selection. In this paper we present our evaluation in the context of texture retrieval. We compare the sparseness-based method with the dispersal-based method, also proposed by Le Borgne et al., and the clustering-based method previously proposed by us. We show that the sparse filters and highly dispersed filters are quite different. In fact we show that highly dispersed filters tend to have lower sparseness. We also show that the sparse filters give better results compared to the highly dispersed filters when applied to texture retrieval. However the sparseness measures are calculated over filter response energies, making this method susceptible to choosing a redundant filter set. This issue is demonstrated and we show that ICF selected using our clustering-based method, which chooses a filter set with much lower redundancy, outperforms the sparse filters.