Unconstrained handwritten numeral recognition with improved radial basis neural network

This paper provides an improved radial basis function neural network (IRBFNN) and applies it in handwritten numeral recognition, On the basis of fuzzy C-means (FCM) and vector quantization (VQ), semi-fuzzy vector quantization (SFVQ) is achieved to obtain the more suitable network structure and higher learning capability. The multiscale compensation algorithm (MSC), which embeds some new small nodes into the original RBFNN to represent the details of the training set, is utilized to improve the normal RBFNN to gain a better recognition accuracy and retain the high generalization. A series of experiments shows that IRBFNN has satisfying performances in the NIST library and some practical handwritten numeral sets.