Hand written digit recognition using BKS combination of neural network classifiers

Abstract

The problem of recognition of handwritten segmented digits irrespective of their size or stroke width is considered. A new approach of combining several different multi-layer perceptron (MLP) neural network classifiers operating on the same image is developed. The classification decisions made by individual MLPs are combined through a method called "behavior-knowledge space" (BKS). The BKS method relies on the behavior of the classifiers on the training set. The pseudo-Zernike moments extracted from the normalized and thinned image of the digit within its bounding circle are used as features. The approach is tested on 3000 digits using three classifiers and a hard error rate of 1.37% is obtained. This is a reduction of almost 50% compared to a single MLP network classifier. The results are also compared to an alternative method of combining the classifiers.<>