A postprocessing method was proposed in this thesis for recognition of handwritten Hangul using pair-wise grapheme discrimination. The proposed method performs a character classification at grapheme level, because most of Hangul recognition errors have been induced by confusion of just one or two component graphemes. The postprocessing process is initiated when a recognition result, which is given by the baseline Hangul recognition system, belongs to a predefined list of frequently confusing pairs. In order to discriminate a character from its confusing counterpart, the proposed method sets up a region within which grapheme comparison is performed. If information in the region is sufficient to discriminate the two characters, a predefined pair-wise discriminator calls upon to verify the result of the baseline system. Otherwise, the region is extended to obtain sufficient information. The sufficiency is determined by the ratio of the common part of the segmentations of the two characters. Several types of pair-wise grapheme discriminators are utilized because a set of features can not discriminate all of the confusion pairs. Three discriminators are used which are image-based, feature point-based and rule-based.
Effectiveness of the proposed method was confirmed by an experiment. The recognition accuracy was improved to 93.2% from 89.2% of the baseline system, which is 37.1% reduction of total errors.
