Deep Dense Multi-level Feature for Partial High-Resolution Fingerprint Matching

Fandong Zhang,Shiyuan Xin,Jufu Feng

DOI: https://doi.org/10.1109/btas.2017.8272723

2017-01-01

Abstract:Fingerprint sensors on mobile devices commonly have limited area, which results in partial fingerprints. Optical sensor can capture fingerprints at very high resolution (2000ppi) with abundant details like pores, incipients, etc. It is quite crucial to develop effective partial-to-partial high-resolution fingerprint matching algorithms. Existing fingerprint matching methods are mainly minutiae-based, with fusion of different levels of features. Their accuracy degrades significantly in our application due to minutiae insufficiency and detection error. In this paper, we propose a novel representation for partial high-resolution fingerprint, named Deep Dense Multi-level feature (DDM). We train a deep convolutional neural network that can extract discriminative features inside any local fingerprint block with certain size. We find that not only minutiae but most local blocks contain sufficient features. Moreover, we analyze DDM and find that it contains multi-level information. When utilizing DDM for partial-to-partial matching, we first extract features block by block through a fully convolutional network, next match the two sets of features pairwise exhaustively, and then select the bi-directional best matches to compute matching score. Experiments indicate that our method outperforms several state-of-the-art approaches.