A Bayesian network classifier and hierarchical Gabor features for handwritten numeral recognition

We present a method of handwritten numeral recognition, where we introduce hierarchical Gabor features (HGFs) and construct a Bayesian network classifier that encodes the dependence between HGFs. We extract HGFs in such a way that they represent different levels of information which are structured such that the lower the level is, the more localized information they have. At each level, we choose an optimal set of 2-D Gabor filters in the sense that Fisher’s linear discriminant (FLD) measure is maximized and these Gabor filters are used to extract HGFs. We construct a Bayesian network classifier that encodes hierarchical dependence among HGFs. We confirm the useful behavior of our proposed method, comparing it with the naive Bayesian classifier, k-nearest neighbor, and an artificial neural network, in the task of handwritten numeral recognition.