Receptive field neural network with shift tolerant capability for Kanji character recognition

The authors present a model of a receptive field neural network as a discriminator for Kanji character recognition. Double emphases on local features of character images are performed in learning: first, generating receptive fields to capture distinctive structures of the characters, and then emphasizing strong differences of the features in the receptive fields which each detects precise positions of the features by shift tolerant LVQ (learning vector quantization) learning. The model is evaluated on printed Kanji characters with high similarity in each of 893 small categories with an average of 2.8 characters per category. It improves the discrimination accuracy from 98.87% without it to 99.55%, a 60% reduction in errors on a test data set of more than 13 fonts. A large scale neural network of three-stage hierarchical structure for Kanji character recognition is presented whose third stage is constructed by this model. It shows 99.0% accuracy on recognition of the multi-font printed 3303 Kanji characters including alphanumeric and symbol on 12000 randomly selected characters from the test data set