Automatic character recognition is necessary for better human computer interaction. Because of the noise and the shape deformation due to the sensor operation, it is not simple to build a practical Hangul recognition system even for printed characters. Neural network or artificial neural system has been reported to be adquate for pattern recognition tasks. However, the number of classes used for the experiments was too small to show the capabilities for the real world problems. Since modern Korean language has thousands of syllables for every day use, Hangul character recognition is a typical problem of thousand classes and it should be interested to know whether neural network approach is capable to cope with the problem.
In this thesis, neural network approach is applied to printed Hangul character recognition with a hope that the problems of conventional pattern recognition techniques are overcomed. Twelve network operate simultaneously. Each network is trained with one of distinct subsets of Hangul characters grouped by the structural similarity. When recognizing an unknown character, each network outputs candidate code and its corresponding matching score. The network with highest matching score wins and its output is selected for the code of the input character. Each network is structured to have one hidden layer and the back-propagation learning algorithm is utilized for updating the forward connection weights. The network also contains backward connections with which the matching score is computed.
In a test case with most frequently used 597 printed Korean characters in Myungjo font, the system achieved 83.6\% correct character recognition rate. However, each network recognized 96\% of characters correctly. Therefore, with an approprieate type classification technique, the recognition rate can be improved about 10\%. Based on this evaluation, we conclude that the neural network approach is adquate for printed Hangul character recognition.