Handwritten Character Recognition Using a MLP
Filippo Sorbello, G. A. M. Gioiello, Salvatore Vitabile · 2020
This chapter describes some algorithms and techniques intended for the handwritten character recognition task using a multilayer perceptron (MLP) neural network. In particular, MLP are widely used for classification problems. The capability to approximate boundary surfaces of arbitrary complexity makes the MLP classifiers universal. The chapter presents a methodology based on a set of appropriate choices for the handwritten character recognition task using a MLP. The use of simple preprocessing techniques, the enhancement of the training set, the choice of opportune activation functions combined with a two step learning process, a learning procedure based on M. J. D. Powell’s conjugate gradient optimization algorithm, and a triple presentation of the test samples are the main ingredients of a recipe that yields results very interesting in terms of both recognition rate and speed. The choice of the activation functions for the hidden and output layers is very critical to reach a good classification rate in each task.