A classifier-based fusion algorithm for latent fingerprint identification based on a neural network

Abstract

Human beings present a singular skin on the surface of their fingers with furrows, ridges, and sweet pores. Ridges and furrows describe distinct forms, such as points with maximum curvature, bifurcation and interruption, and particular ridges’ contours. Experts have classified those forms as level 1, 2, or 3 features. When the finger surface touches an object, it prints the features in a 2D image termed fingerprint due to the grease and the sweat released by the pores. Hitherto, there is no report of two fingerprints having the same features, not even in identical tweens. Fingerprints acquired from the object surfaces with unknown identity are latent fingerprints. Latent fingerprints have practical applications in criminal investigations and justice administration. Such sensible applications demand high accuracy and speed in the identification of a latent fingerprint. Although some authors have proposed latent fingerprint identification algorithms, they still consider insufficient the achieved identification rates for satisfying the sensitivity of the latent fingerprint applications. Some enhancements to the identification of latent fingerprints have been reported with fusion algorithms. However, the typical fusion scheme has been focused on weighted sums with weights empirically determined. The trending use of weighted sums has left room for improvements using classifier-based fusions. Additionally, the literature related to the latent fingerprint identification lacks of an exhaustive analysis of the suitability of the multiple fingerprint feature representations proposed, and the quantification of such a suitability. In this research, we analyze the appropriateness of several fingerprint feature representations for representing latent fingerprints; and we have found a preference for minutia descriptors. Hence, we develop a protocol for evaluating minutia descriptors in a closedset identification. With such a protocol, we determine the merit of nine minutia descriptors suitable for identifying latent fingerprints. As a result, we select four minutia descriptors as candidates for a fusion algorithm and tune their parameters for latent fingerprint identification. Next, we evaluate the four minutia descriptors with their global matching algorithms on subsets of latent fingerprints with good, bad, and ugly qualities. We find two of them reaching the highest identification rates for all subsets and ranks. Therefore, we propose a latent fingerprint identification algorithm with the fusion of these two algorithms using a neural network and four attributes as input, which characterize the fingerprints’ similarity. Experiments show that our proposal improves the baseline algorithms in 13 of 15 datasets created with databases NIST SD27, MOLF-IIITD, and GCBD. Our fusion algorithm reports the highest rank-1 identification rate (71.32%), matching the latent fingerprints in the NIST SD27 against 100,000 fingerprints, using only minutiae. Our algorithm takes six milliseconds to compare a fingerprint pair, which is a good time.