On-line Korean character recognition by using two types of neural networks

We present an online handwritten Korean character recognition method which uses 2 different neural networks. By noting the fact that a Korean character is made of two dimensional composition of strokes, the method recognizes a character by identifying strokes and composing them. The first network receives each stroke data and classifies it to one of the predefined stroke classes. The input of this network is the direction feature vector of a stroke. The second network recognizes the character by receiving the class codes of all strokes constituting a character and the information of the relative positions between two consecutive strokes. This network is trained to acknowledge all possible stroke orders and relative positions. It appears that this training can be made by using only a small portion of the entire Korean character set. Experimental results indicate that the method is promising.