Unsupervised Fingerprint Dense Registration

Yuwei Jia, Zhe Cui, Fei Su

Abstract: Fingerprint registration is still a challenging task due to the large variation of fingerprint quality. Meanwhile, existing supervised fingerprint registration methods need sufficient amount of labeled fingerprint pairs which are difficult to obtain. In addition, the training data itself may not include enough variety of fingerprints thus limit such methods’ performance. In this work, we propose an unsupervised end-to-end framework for fingerprint registration which doesn’t require labeled fingerprint data. The proposed network is based on spatial transformer networks, and can be applied flexibly to achieve a better results by being used recursively. Experiment results show that our method gets the state-of-the-art matching scores while preserving the good ridge structure of fingerprints, and achieves competitive matching accuracy through score fusion when compared with supervised methods.