The difficulties of hand-written character recognition are mainly due to the variations on writing style. Hand-written Hangul recognition is more difficult because of Hangul's huge character set. Fortunately, Hangul characters are composed of basic alphabets and, therefore, structural analysis attempting to recognize component alphabets is believed the way to pursue.
In this thesis, a system of neural networks is developed to recognize hand-written Hangul by stroke analysis. From the strokes extracted from training data, a feature map is constructed by Kohonen's learning mechanism. In the recognition phase, input strokes are encoded in terms of the features. A neural network, then, takes a group of encoded strokes as input and classifies it into one of the six Hangul types. Finally a selected neural network produces its character code.
An experiment is conducted with five sets of a single writer's 522 hand-written Hangul characters which cover 98\% of daily use. The first three sets are used for training and the last two sets are used for testing. After a training, the system achieved 71\% recognition rate for the test sets, while 96\% recognition rate for the trained sets. Although an extensive training may improve the recognition rate, the system may not be able to overcome the excessive variations in hand-writen characters.