An MLP Based Approach for Computing the Shape of a Pattern Class
Deba Prasad Mandal · IETE Journal of Research · 1996
We have proposed here a multilayer perceptron (MLP) model based approach for computing the shape of a pattern class. The problem is viewed here as a two class classification problem where the first class consists of the training samples of the pattern class and the other class, called a reference set, consists of some equidistant points in the feature space. A modification of the MLP model using back propagation (BP) algorithm is carried out to formulate a rough shape determining algorithm. The shape of a pattern class is obtained by passing through two different stages. In the first stage, the external boundary is calculated whereas the second stage finds the internal boundaries (holes) of the pattern class. In both the stages, the rough shape determining algorithm is initially applied, and then the rough estimated shapes are improved by the usual BP algorithm. The external and internal shapes, as obtained in the previous two stages, are finally combined to decide about the complete shape of the pattern class. Various synthetic data sets are considered to demonstrate the effectiveness of the proposed procedure. We have also verified the convergence (with sample size) of the estimated shapes to the original pattern class.