Pattern recognition by topology free spatio-temporal feature map

Abstract

This paper introduces a novel concept that eliminates the need for a topologically ordered feature map required for a pattern classification task. Spatio-temporal feature maps (extensions to Kohonen's self-organizing feature maps) have been shown earlier to provide enhanced classification performances over self-organizing feature maps. However, a topologically ordered feature map is still required as a basis to form a spatio-temporal map. In this paper, it is shown that by picking suitable samples of the input patterns as weights and ensuring that the selected weights are stratified and contained within the convex hull of the input space, the high classification performance of the spatio-temporal feature maps can still be retained. Such a formation of spatio-temporal feature map has no relation to topology preservation concept and the new classification paradigm is, therefore, topology free. The simulation results on 8-class texture and 12-class 3D object feature data sets confirm the high classification performance without the need for computationally expensive training required to obtain topologically ordered feature maps.