Hierarchical Bayesian Network for Handwritten Digit Recognition

This paper introduces a hierarchical Gabor features(HGFs) and hierarchical bayesian network(HBN) for handwritten digit recognition. The HGFs represent a different level of information which is structured such that the higher the level, the more global information they represent, and the lower the level, the more localized information they represent. The HGFs are extracted by the Gabor filters selected using a discriminant measure. The HBN is a statistical model to represent a joint probability which encodes hierarchical dependencies among the HGFs. We simulated our method about a handwritten digit data set for recognition and compared it with the naive bayesian classifier, the backpropagation neural network and the k-nearest neighbor classifier. The efficiency of our proposed method was shown in that our method outperformed all other methods in the experiments.