Latent Fingerprint Matching via Dense Minutia Descriptor

TL;DR

This work tackles the challenging problem of latent fingerprint matching by introducing Dense Minutia Descriptor (DMD), a three-dimensional, dual-branch neural descriptor that jointly models texture and minutiae information while using a segmentation map to constrain matching to foreground regions. The method aligns patches to anchor minutiae, employs a 2D positional embedding, and normalizes matching scores based on overlap, enabling robust comparison despite partial and degraded latent prints. Training combines CosFace-based losses for both descriptor types, a segmentation and minutiae regression loss, and a similarity loss between plain and distorted fingerprints, with sophisticated minutiae sampling and data augmentation to simulate distortions. Empirical results on NIST SD27 and N2N Latent demonstrate state-of-the-art performance, and a binary variant further improves efficiency while preserving accuracy, signaling strong potential for scalable latent fingerprint indexing and secure template applications.

Abstract

Latent fingerprint matching is a daunting task, primarily due to the poor quality of latent fingerprints. In this study, we propose a deep-learning based dense minutia descriptor (DMD) for latent fingerprint matching. A DMD is obtained by extracting the fingerprint patch aligned by its central minutia, capturing detailed minutia information and texture information. Our dense descriptor takes the form of a three-dimensional representation, with two dimensions associated with the original image plane and the other dimension representing the abstract features. Additionally, the extraction process outputs the fingerprint segmentation map, ensuring that the descriptor is only valid in the foreground region. The matching between two descriptors occurs in their overlapping regions, with a score normalization strategy to reduce the impact brought by the differences outside the valid area. Our descriptor achieves state-of-the-art performance on several latent fingerprint datasets. Overall, our DMD is more representative and interpretable compared to previous methods.

Figures (6)

• Figure 1: Compared with (a) one-dimensional minutia descriptor, (b) our Dense Minutia Descriptor (DMD) is a three-dimensional representation, and explicitly considers the overlapping area for score normalization. Score normalization is denoted as *.
• Figure 2: The detailed structure of our DMD extraction network. The content boxes display operation names, output channels, and spatial scales separated by commas. The third one is omitted if scale equals 1.
• Figure 3: The process of selecting training minutiae pairs.
• Figure 4: Latent fingerprint matching performance on NIST SD27 (a) (c) and N2N Latent (b) (d).
• Figure 5: Descriptor visualization on different patch images. Descriptors of MinNet are resized to three-dimension form for visualization. We select a specific channel from the aforementioned descriptors and convert it to a binary format to enhance visualization.