Design and Evaluation of a Hierarchical Architecture for Handwritten Character Recognition

Handwritten character recognition represents a problem that was approached in many ways by the scientists. Although a generally solution for any type of handwritten characters not founded yet, the obtained results give hopes to continue the researches in this field. In this paper, we present a software architecture used in character recognition: Hierarchical Neural Network (HNN) architecture (Halici et al. 1999). Using this architecture, we have approached two important aspects from the handwritten character recognition task: input data preprocessing and clustering. The input data preprocessing consists of a compression applied to the input character in order to obtain a reduced standard character matrix. The reduced character matrix will be used, in an efficient manner, as input data, in the character recognition application. The first stage in the character recognition is the clustering stage performed by a Self-Organizing Feature Map (SOFM) Neural Network (NN). The clustering process consists in fact in a classification based on the input vectors similarities.