Representation and parameter estimation of hierarchical random graph and its application to handwritten Hangul recognition

Abstract

In this thesis, a hierarchical random graph (HRG), which is a stochastic graph for handwritten character modeling, is proposed. In addition, based on the HRG, a handwritten Hangul recognition system also has been developed. In the HRG, the bottom layer is constructed with chain graphs to describe various strokes, while the next upper layers are constructed with random graphs [1] to model spatial and structural relationships between strokes and between sub-characters. Since the proposed HRG is a stochastic model, the recognition can be formulated into the problem that chooses a model producing maximum probability given an input data. In this context, matching score is obtained not by any heuristic similarity function, but by the observation probability calculated by multiplying observation probabilities of features of an input character. To estimate the model parameters of the HRG, we apply embedded training, which estimates all the models in the HRG at the same time so as to avoid manual segmentation of an input character. As the criteria of parameter estimation in embedded training, we adopt maximum likelihood estimation (MLE) and maximum mutual information estimation (MMIE), which are representative methods to estimate the model parameters in statistical approaches, especially having been used for the parameter estimation of hidden Markov models (HMM). The criteria are examined in the information theoretic view, and a new criterion based on minimum entropy estimation (MEE) is proposed. Since it is not necessary to assume that the parameter space of models is correct, MEE will perform not less than any other estimation methods under the condition that the training data size is large enough. Some experiments with handwritten Hangul characters show the performance of the three estimation methods. Besides, many issues for automatic parameter estimation of the HRG are discussed and implemented, including exclusive training, parameter smoothing, and initial parame...