Hierarchically structured neural networks for printed Hangul character recognition

In this paper, we propose a hierarchical neural network which practically recognizes printed Hangul(Korean) characters. This system is composed of a type classification network and six recognition networks. The former classijies input character images into one of the six types by their overall structure, and the latter further classify them into character code. Furthermore, a training scheme including systematic noises is introduced for the improving the generalization capabilities of the networkx Experiments are conducted with most frequently used 990 printed Hanqul characters. By the noise included training, the recognition rate amounts up to 98.282, which is superior to the conventional backpropagation network. Afer all, neural network approach turns out to be very reasonable through a comparison with statistical classifer and an analysis of generalization capability. I