Abstract
Contactless fingerprint identification has emerged as a reliable and user-friendly alternative for personal identification in a range of e-business and law-enforcement applications. It is however quite known from the literature that the contactless fingerprint images deliver remarkably low matching accuracies as compared with those obtained from the contact-based fingerprint sensors. This paper develops a new approach to significantly improve contactless fingerprint matching capabilities available today. We systematically analyze the extent of complimentary ridge-valley information and introduce new approaches to achieve significantly higher matching accuracy over state-of-art fingerprint matchers commonly employed today. We also investigate least explored options for the fingerprint color-space conversions, which can play a key-role for more accurate contactless fingerprint matching. This paper presents experimental results from different publicly available contactless fingerprint databases using NBIS, MCC and COTS matchers. Our consistently outperforming results validate the effectiveness of the proposed approach for more accurate contactless fingerprint identification.
Introduction
Traditional fingerprint sensors require physical contact, which raises concerns about hygiene (e.g., spreading viruses like SARS), sensor wear, and "latent" prints that could be lifted for spoofing attacks. Contactless systems—using smartphone cameras or specialized setups—solve these issues but introduce a new challenge: significantly lower matching accuracy compared to traditional "touch" sensors.
In a contact sensor, ridges touch the glass and appear dark, while valleys do not and appear bright. In contactless images, however, the 3D structure of the finger and ambient lighting cause "polarity reversal." Depending on the angle of light, a ridge might appear bright in one spot and dark in another. Standard algorithms, designed for the consistent "dark ridge/bright valley" paradigm, struggle to find reliable features (minutiae) in these inconsistent images.
Methods
The proposed "collaborative approach" is based on two main pillars:
1. Ridge-Valley Fusion
Instead of just processing the raw image, the researchers create an inverted version of the fingerprint. They extract minutiae from both. In the inverted image, the "valleys" of the finger effectively become "ridges." By matching both sets of features and fusing the resulting scores, they capture complementary information that a single pass would miss.
2. Luma Grayscale Transformation
Most systems convert color images to grayscale using a simple linear combination of RGB channels. This study investigated Luma representation, which uses non-linear gamma-corrected versions of the color channels ($I_{Luma} = 0.2126 R' + 0.7152 G' + 0.0722 B'$). This method preserves higher contrast, making it easier for algorithms to distinguish fine ridge details and suppress "spurious" (fake) minutiae.
Datasets and Evaluation
The study evaluated the approach on two major public datasets:
- PolyU Contactless to Contact-based Database: 2,016 contactless 2D fingerprints from 336 subjects.
- Benchmark 2D/3D Fingerprint Database: 2,000 images from 1,000 different fingers, featuring higher pose variations and distances from the sensor.
Testing was performed using three different matchers: the open-source NBIS (NIST), the MCC (Minutiae Cylinder Code), and a commercial off-the-shelf (COTS) tool, VeriFinger.
Key Insights and Ideas
- Valleys are Hidden Assets: In contact-based sensing, valley information is essentially lost background. In contactless sensing, valleys are fully visible. By inverting the image, the system can detect "ridge endings" that were actually "valley bifurcations," providing a second, independent map of the finger.
- Polarity is Fluid: Because of the finger's 3D curve, a single ridge can have different polarities across its length. Neither the raw nor the inverted image is "correct" on its own; rather, they are both partially correct, and their combination is more robust.
- Color Conversion Matters: The move from "Ordinary" grayscale to "Luma" grayscale significantly reduced the Equal Error Rate (EER) by nearly 50% in some tests because it better handled the way human skin reflects light.
Key Results
The improvements were consistent across all matchers and databases:
- Accuracy Boost: On the PolyU database using the COTS matcher, the EER dropped from 0.81% (raw) to 0.26% (combined ridge-valley), a 67.9% improvement.
- Luma Superiority: Using Luma grayscale with the NBIS matcher improved the EER by roughly 50% compared to ordinary grayscale.
- Robustness: Even when the inverted image performed worse on its own than the raw image, the combination of the two almost always outperformed the raw image alone, proving the "collaborative" nature of the features.
Conclusion and Future Work
While the results are promising, challenges remain regarding pose and distance variations. Future work should involve deep learning frameworks like Faster-RCNN to automatically locate and crop the finger in real-time from complex backgrounds. Additionally, the computational overhead of processing images twice (raw and inverted) needs optimization for deployment on mobile devices.
Summary updated