Learning Optimization in a MLP Neural Network Applied to OCR

This paper focuses on the possibilities of optimization of the training process of an MLP neural net using Backpropagation as a learning algorithm, employed as a classifier in an Optical Character Recognition (OCR) application. Also, the process for determination of a set of optimal parameters describing the characters that conform each class is described. The processing and analysis of the images in BMP, GIF, JPG and TIF format are included. A comparative study of the possibilities of improvement of the learning process of an MLP net employing heuristics for its design and training is made. As a fundamental result a substantial improvement of the net learning process is obtained and an OCR of great reliability is built.