A lot of research is being done on pairwise discrimination of handwritten characters, like Chinese characters and English. To do that, features for pairwise discrimination must be selected. However, most of existing works select features manually, or use features that are used for implementing the baseline recognizer.
In this thesis, we propose a pairwise discrimination method of handwritten Hangul characters based on discriminative feature selection. First, a large size of candidate feature set is constructed per each character by using popular features in the offline character recognition literature. Each feature is represented by a probability mass function whose parameters are estimated from the training data. Second, features in the feature set are selected in the objective of discriminating characters, which are measured by the Bayesian decision error. Last, a pairwise discriminator is constructed by combining the selected features based on the $na"{I}$ ve Bayesian rule.
The proposed pairwise discrimination method is applied to the recognition of handwritten Hangul characters. The experimental result with 489 characters from postal mail envelops showed the effectiveness of the proposed method; the recognition accuracy was improved to 81.7% from 80.9% of the baseline system (4.6% relative error reduction).
