Salima Nebti, Abd Ellah Boukerram
Departement of computer science. Ferhat Abbess University 19000 Sétif. Algeria.

Abstract

The work below describes a particle swarm optimization (PSO) based approach for learning a multi layer perceptron network in order to solve the handwritten digits recognition problem. Actually, the gradient descent methods are widely used for this aim, since they can provide good results in a reduced time. Unfortunately, they can converge towards local optima. Within this issue, the present paper outlines the hybridization of PSO and neural networks to deal with digits recognition through the use of PSO as an initialization step for gradient backpropagation. The proposed approach has been applied to a variety of handwritten digits with different characteristic features namely the principal component analysis and Gabor features, good quality results have been obtained when compared to a pure PSO based learning method due to the hybridization of global search using PSO and local search using the gradient back-propagation.

Keywords: handwritten digits recognition, artificial neural networks, backpropagation, features extraction.