Extraction of Binary Features from Fingerprint Topology

Abstract

We propose a technique of extraction repeatable binary string from fingerprint images. We extract binary features from topological relations between minutiae points. For an arbitrary minutiae point we trace neighboring ridges until we encounter event: minutiae or projection of minutiae. Then we encode these events. Thus we obtain 50-100 bit descriptions for each minutiae point. In order to extract longer binary string, we propose two techniques. First, we select 2^l binary strings (centers) of the same length as topological descriptor. Then for each minutiae descriptor we select k nearest centers. Selected subset of centers is represented by 2^l bit long binary string. The self-aligned technique has significant errors (FRR=30% at FAR=0.1%). The second tehnique requires public helper. We concatenate descriptions of 5-8 minutiae points. Data that are necessary to find and order minutiae subset is stored in a public helper. Thus we extracted 384-512 bit binary string with approximately 20% erroneous bits. Errors are corrected using two-layer BCH-major voting codes. Experiments on FVC2002DB1 dataset and private dataset (100 different fingers, with 3 samples per finger) show that 20-40 bit error-free binary string can be reproduced from genuine fingerprint with 90% success rate.