Weakly Supervised Fingerprint Pore Extraction With Convolutional Neural Network

Abstract

Fingerprint recognition has been used for person identification for centuries, and fingerprint features are divided into three levels. The level 3 feature is the fingerprint pore, which can be used to improve the performance of the automatic fingerprint recognition performance and to prevent spoofing in high-resolution fingerprints. Therefore, the accurate extraction of fingerprint pores is quite important. With the development of convolutional neural networks (CNNs), researchers have made great progress in fingerprint feature extraction. However, these supervised-based methods require manually labelled pores to train the network, and labelling pores is very tedious and time consuming because there are hundreds of pores in one fingerprint. In this paper, we design a weakly supervised pore extraction method that avoids manual label processing and trains the network with a noisy label. This method can achieve results comparable with a supervised CNN-based method.