An on-line character recognition system for handwritten alphanumerics using the backpropagation learning algorithm has been developed. This system accepts sequences of strokes from transparent degitizing device and segments them into sequence of characters. Then recognizer is called to produce character code. Input characters are assumed to be run-on characters, which means that characters can be written in one or more strokes but a stroke cannot span over more than one character. This system is designed to be user independant, stroke order and number free.
Character segmentation as well as recognition is performed by neural networks. Features of single stroke as well as relationships among consecutive strokes are used to make stroke grouping decisions. The decision process is trained by the backpropagation algorithm. In the recognition process, many on-line features are extracted and used. This training is also performed by the backpropagation.
Experiments are performed for writings of 23 different writers. 2 sets of each writers are used for training and the remaining set is used for testing. The character recognition rate is about 94%. However, the recognition rate is over 95.4% if we exclude the errors unavoidable by shape analysis.