Region-mapping neural network model for pattern recognition

Abstract

In general, the process for multilayer feedforward neural network in pattern recognition is composed of two phases: training and classifying. The aim of the training phase is to make the network output meet the desired output given by the training patterns as possible. It demands a map of point to point, which is so strict that it often causes the criterion inconsistence between training and classifying. Consequently the recognition rate would be decreased. Region-mapping model has changed the output space from one point to a certain supervisor region so that it has overcome the shortcoming of inconsistent problem between training and testing as common multilayer perceptron (MLP) does. Furthermore, it has saved much computing time by mapping the input data to an output area rather than an output point. This paper presents a Region-mapping model with quarter hyper globe as supervisor region. The gradient decent algorithm is applied to this model. In order to illustrate the effect of our propounded model, a hand-written letter recognition problem is put into experiment. Moment invariant features are used as input parameters. The simulation results show that the region-mapping model has much better characteristics than those common multiplayer perceptrons. Also, the quarter hyper globe rule is more reasonable than the hypercube one.