Utilização de redes bayesianas como agrupador de classificadores locais e global.

Abstract

The classification problem in pattern recognition can be viewed as a probability distribution estimation task. Recent developments try t o model it as a weight sum of distributions which is a parametric approach, since weights and parameters should be estimated. In this work the target distribution is reached without the need to estimate parameters from a model distribution. Considering that the output of classifiers are probability measurements, a Bayesian network is used t o combine local and global classifiers. Briefly, the main objective of this work is to present a methodology that establishes how t o partition the feature space in order to generate a set of classifiers and group them in a framework that combines their outputs. A case study was developed for a handwritten digit recognition application. The results reveal that the proposed system is competitive with the best classifiers pointed in the literature. vi