DIE - Dipartimento di Ingegneria Elettrica, Universita di Palermo,
Viale delle Scienze, 90128 Palermo, Italy
E-Mail: gioiello@diepa.unipa.it
E-Mail: sorbello@diepa.unipa.it
E-Mail: tarantino@diepa.unipa.it
CRES - Centro per la Ricerce Elettronica in Sicilia,
Via Regione Siciliana 49, 90046 Monreale, Italy
E-Mail: vassallo@ipacres.cres.it
Abstract
The aim of this work is the derivation of simple algorithms and techniques for using an MLP neural network (MultiLayer Perceptron) for a handwritten character recognition task in an efficient way. A set of techniques including the optimization algorithm used for the learning, the activation functions chosen for the hidden and output layer, the preprocessing applied, artificial enhancement of the training set and multipresentation of the test samples, lead to high performance in terms of recognition rate, preprocessing and classification speed. The simplicity of the techniques used also makes possible an implementation on a traditional serial computer with reasonable speed. A digital implementation is also outlined.