Chinese character recognition with neural nets classifier

An optical Chinese character recognition system using a neural nets classifier is presented. To extract stable information and reduce the effect of varying stroke widths, a feature extraction scheme which retrieves character boundaries and then quantizes the pixels to four possible orientations is suggested. To improve the learning speed and to reduce the architectural complexity, a perceptron with no hidden layer is adopted. In the learning phase, the link weights of the perceptron are adjusted iteratively by a back propagation learning algorithm. From simulation results, the recognition rates are 91% and 99% for handprinted and multifont Chinese characters, respectively. The rates are significantly superior to those obtained with a traditional nearest-mean classifier