In this thesis, we propose a hybrid recognition method, and show its usefulness in recognizing on-line cursive Hangul (the Korean writing system) characters. Although many successful handwriting recognizers using statistical method or structural method have been proposed, they could not overcome the limitation of single method. Since the advantage of one method appears to be the disadvantage of the other, many researchers have tried to make a hybrid model. The goal of this thesis is to design an effective framework combining the statistical method and the structural method that incorporates the advantages of the individual strengths of both.
A stochastic attributed network is constructed to combine the statistical method and the structural method. The architecture consists of three layers of recognition scheme. The network in the baseline layer has a finite number of nodes and arcs representing the rules of character composition from graphemes. Each arc of baseline layer expands into a set of statistical models, in particular hidden Markov models in the statistical layer, and each node makes use of structural knowledge sources in the structural layer. The statistical recognizer produces intermediate recognition results including the label and the boundaries of grapheme. The intermediate results are passed to structural layer through baseline layer, then the structural recognizer analyze them. Shape analyzer representing global shape of grapheme, position analyzer representing relative position in a character, and pairwise discriminator representing discriminative features of confusing pair are used in the structural layer.
Consequently, recognition of handwritings in the proposed method corresponds to finding the most probable path through the stochastic attributed network. Viterbi algorithm has been used, however, first-order Markov assumption and observation independence assumption applied to reduce complexity seem to create some errors due to missing globa...