Fingerprint Classification Based on Curvature Sampling and RBF Neural Networks

Abstract

This paper presents new five-class fingerprint classification algorithms based on combination of curvature sampling and radial basis function neural networks (RBFNNs). The novel curvature sampling algorithm is proposed to represent tendencies and distributions of ridges’ directional changes with 25 sampled curvature values. The normalized and organized curvature data set is as input feature vector for RBFNNs and the output is formed result. The probability density is defined to describe the clustering ability of an input vector and used to select hidden layer neurons adaptively. The algorithms are validated in fingerprint databases NIST-4 and CQUOP-FINGER, the best classification accuracy is 91.79% at 20% rejection rate. It shows good balance for classification of arch and tented arch types and it needn’t detect singular points. The result indicates that this algorithm can satisfy the requirement of fingerprint classification well and provides a new and promising approach.