Optical Chinese character recognition using probabilistic neural networks

Building on previous work in Chinese character recognition, we describe an advanced system of classification using probabilistic neural networks. Training of the classifier starts with the use of distortion modeled characters from four fonts. Statistical measures are taken on a set of features computed from the distorted character. Based on these measures, the space of feature vectors is transformed to the optimal discriminant space for a nearest neighbor classifier. In the discriminant space, a probabilistic neural network classifier is trained. For classification, we present some modifications to the standard approach implied by the probabilistic neural network structure which yield significant speed improvements. We then compare this approach to using discriminant analysis and Geva and Sitte’s Decision Surface Mapping classifiers. All methods are tested using 39,644 characters in three different fonts.