Hand Written Character Feature Extraction Using Non-Linear Feedforward Neural Networks

Abstract

In this paper, an artificial neural network is proposed for feature extraction of hand written characters. The learning algorithm is developed based on a proposed modified Sammon's stress for our feedforward neural networks, which can not only minimize intra class pattern distances but also preserve interclass distances in the output feature space. The proposed feature extraction method tries to calculate rough classes using a Competitive Learning neural network, which is an unsupervised neural network. Then the proposed neural network was used with modified Sammon's stress to perform feature extraction using information obtained by means of a Competitive Learning Network. The features thus obtained were compared with a standard PCA neural network and a neural network using Sammon's stress in terms of their classification accuracy. Two numerical criteria were used for performance evaluation of the features-the normalized classification error rate and modified Sammon's stress. It is found that proposed modified Sammon's stress provides features that are more efficient based on these two numerical criteria.