Abstract:
In this paper, we propose a novel approach to enhance the classification performance of Self-Organizing Map based classifiers by extending the input pattern dimensionality. Depending on the technique used in projecting a pattern onto a high dimensional manifold, it is shown that the input patterns that lie very close to each other can be separated. Such projections alters the probability distribution in feature space and hence the topological ordering in SOM as well. The technique is demonstrated on a set of 8-dimensional feature data obtained from 8-class texture images. Kohonen's SOM and its extensions to higher-order correlation maps are tested with the proposed technique and the results indicate a consistent improved performance for a specific type of projection.