Latent Fingerprint Registration via Matching Densely Sampled Points

Shan Gu, Jianjiang Feng, Jiwen Lu, Jie Zhou

Abstract

Latent fingerprint matching is a very important but unsolved problem. As a key step of fingerprint matching, fingerprint registration has a great impact on the recognition performance. Existing latent fingerprint registration approaches are mainly based on establishing correspondences between minutiae, and hence will certainly fail when there are no sufficient number of extracted minutiae due to small fingerprint area or poor image quality. Minutiae extraction has become the bottleneck of latent fingerprint registration. In this paper, we propose a non-minutia latent fingerprint registration method which estimates the spatial transformation between a pair of fingerprints through a dense fingerprint patch alignment and matching procedure. Given a pair of fingerprints to match, we bypass the minutiae extraction step and take uniformly sampled points as key points. Then the proposed patch alignment and matching algorithm compares all pairs of sampling points and produces their similarities along with alignment parameters. Finally, a set of consistent correspondences are found by spectral clustering. Extensive experiments on NIST27 database and MOLF database show that the proposed method achieves the state-of-the-art registration performance, especially under challenging conditions.

Introduction: The Minutiae Bottleneck

In forensic science, latent fingerprints—those left unintentionally at crime scenes—are often smudged, partial, or overlaid with background noise. Traditional automated fingerprint identification systems (AFIS) rely heavily on minutiae, specific points where ridges end or split. However, when a latent print is of poor quality or covers a small area, extracting a sufficient number of reliable minutiae becomes a significant bottleneck. If registration (alignment) fails due to a lack of landmarks, the subsequent matching process inevitably fails.

Comparison of minutiae extraction versus the proposed dense matching approach. (a) Manually marked minutiae, (b) minutiae estimated by FingerNet, and (c) the proposed registration result.

To address this, we propose a non-minutia latent fingerprint registration method. Instead of relying on sparse landmarks, our approach estimates spatial transformation through a dense fingerprint patch alignment and matching procedure, effectively utilizing the entire available ridge pattern.

Methodology

The proposed framework follows a coarse-to-fine registration strategy that bypasses explicit minutiae extraction or orientation field estimation for the latent print.

Dense Sampling and Local Alignment

Rather than searching for specific landmarks, the algorithm places a uniform grid of sampling points across the fingerprint. For every pair of sampling points between the latent and reference prints, a Siamese network (based on VGG-13) predicts the relative rotation and translation required to align the local image patches. This relative transformation estimation is a key innovation; it is more robust than traditional methods that attempt to calculate an absolute orientation for noisy latent patches.

Patch Matching and Global Consistency

Once patches are aligned, a second Siamese network extracts deep descriptors to determine similarity. To filter out false correspondences and ensure geometric consistency, the system employs spectral clustering. This step ensures that the final set of matching points forms a coherent global transformation across the entire image.

Training with Synthetic Data

Due to the scarcity of labeled latent fingerprint datasets, the model was trained using a massive synthetic dataset. The researchers utilized CycleGAN and Thin Plate Splines (TPS) to simulate realistic latent prints from high-quality rolled prints, capturing the specific noise, pressure, and blur characteristics typical of forensic evidence.

Experimental Results

The method was evaluated on the NIST27 and MOLF databases. On NIST27, the "Proposed-precise" version improved Rank-1 identification accuracy from 61.6% to 70.1% compared to previous state-of-the-art methods.

A detailed view of the dense correspondences established between a latent fingerprint and a rolled reference print.

The most significant performance gains were observed in the "Ugly" category of NIST27, where image quality is poorest. In these cases, registration accuracy improved by over 12% compared to minutiae-based approaches. Furthermore, the model demonstrated strong generalization on the MOLF database, increasing Rank-1 accuracy from 15.2% to 19.8%.

Conclusion

By shifting the focus from sparse minutiae to dense patch matching, this research provides a robust solution for the registration of low-quality latent fingerprints. While the current implementation faces computational challenges—taking approximately 3.49 seconds per pair—it represents a significant step forward for forensic identification, particularly for the most challenging "ugly" prints where traditional systems fail. Future work will focus on optimizing the system for large-scale database searches and improving handling of extreme skin distortion.

Summary updated

Abstract

Latent fingerprint matching is a very important but unsolved problem. As a key step of fingerprint matching, fingerprint registration has a great impact on the recognition performance. Existing latent fingerprint registration approaches are mainly based on establishing correspondences between minutiae, and hence will certainly fail when there are no sufficient number of extracted minutiae due to small fingerprint area or poor image quality. Minutiae extraction has become the bottleneck of latent fingerprint registration. In this paper, we propose a non-minutia latent fingerprint registration method which estimates the spatial transformation between a pair of fingerprints through a dense fingerprint patch alignment and matching procedure. Given a pair of fingerprints to match, we bypass the minutiae extraction step and take uniformly sampled points as key points. Then the proposed patch alignment and matching algorithm compares all pairs of sampling points and produces their similarities along with alignment parameters. Finally, a set of consistent correspondences are found by spectral clustering. Extensive experiments on NIST27 database and MOLF database show that the proposed method achieves the state-of-the-art registration performance, especially under challenging conditions.
BibTeX
@article{gu2021latent,
  author = {Shan Gu and Jianjiang Feng and Jiwen Lu and Jie Zhou},
  title = {{Latent Fingerprint Registration via Matching Densely Sampled Points}},
  journal = {IEEE Trans. Inf. Forensics Secur.},
  year = {2021},
  doi = {10.1109/tifs.2020.3032041},
}
Figure