Pyramid architecture classification tree

Abstract

This paper proposes a novel pattern recognition algorithm-the pyramid architecture classification tree (PACT). The learning phase of the recognition system consists of two steps: a pyramid making step and a decision tree making step; all training patterns are preprocessed by the pyramid structure and the results are used for making a decision tree. PACT directly copes with a bitmap array having the two dimensional topology and needs no feature extraction. For evaluation of the performance of PACT, various experiments using a handprint Japanese character database were carried out. The results show that PACT can realize about 50 times faster training speed than that of conventional decision tree classifiers, and classifies patterns in far higher speed than nearest neighbor matching algorithms.