Deep and discriminative feature learning for fingerprint classification

Fingerprint classification is still a challenging problem because of its large intra-class variance, small inter-class variance and strong noise in the fingerprint patterns. Traditional methods usually utilize some human-defined features to classify fingerprints, however, these features are relatively shallow and local, and they cannot solve this problem well. In this paper, we propose a deep convolutional neural network called Res-FingerNet to solve this problem. The network can extract more abstract and global features from fingerprint images. Moreover, in order to reduce intra-class variance and enlarge inter-class variance of the fingerprints, we utilize center loss in the network training stage so that the learned deep features are more discriminative for fingerprint classification task, and our experimental results show that the classification accuracy increases about 1.5% by center loss. The classification method has been measured on the public fingerprint database NIST-DB4 and it achieves a classification accuracy of 97.9% which surpasses most of existing fingerprint classification methods.