Abstract
Biometric systems relying on a single modality often face limitations in accuracy and robustness. This paper presents a multimodal biometric system that captures both face and iris data from a single high-resolution Near Infrared (NIR) image. By fusing these modalities at the score level, the proposed system achieves significantly higher recognition accuracy and improved resilience to varying image quality compared to unimodal approaches.
Introduction
Traditional unimodal biometric systems, such as those based solely on face or iris recognition, possess inherent vulnerabilities. Face recognition is sensitive to environmental factors like illumination and pose, while iris recognition, though highly accurate, often requires significant user cooperation for high-quality acquisition. Multimodal systems address these issues but frequently introduce hardware complexity by requiring multiple sensors. This research explores a "single-shot" approach using a high-resolution Near Infrared (NIR) sensor to capture both facial and iris features simultaneously, providing a balance between user convenience and security.
Methodology
The proposed system utilizes a 10-megapixel CCD camera equipped with 850nm NIR LEDs and a band-pass filter. This configuration ensures consistent image quality regardless of ambient visible light, making the system invariant to environmental changes.
Data Acquisition and Segmentation
From a single high-resolution capture (2736 × 3648 pixels), the system automatically segments the face and both the left and right irises. The high resolution is a critical engineering requirement; the authors calculate that at least a 9-megapixel sensor is necessary to maintain sufficient iris detail (typically 320×280 pixels) within a full-face shot.
Feature Extraction and Fusion
- Face Recognition: Features are extracted using Principal Component Analysis (PCA), projecting images into a lower-dimensional "Eigenface" space.
- Iris Recognition: The system employs the Daugman algorithm, utilizing 2D Gabor wavelets to generate a 256-byte IrisCode and calculating similarity via Hamming Distance.
- Score-Level Fusion: Matching scores from the face and both irises are normalized using a Min-Max approach. These scores are then combined using Sum or Product rules to produce a final verification decision.
Experimental Results
The system was evaluated on a custom NIR database containing 930 images from 112 subjects. Two test sets were used: a "Closed Set" of known subjects and an "Open Set" designed with lower-quality iris images to test robustness.
The fusion approach showed dramatic improvements, especially in difficult scenarios. In the Closed Set, standalone face recognition achieved 97.35% accuracy at a False Accept Rate (FAR) of 0.001, which rose to 99.75% when fused with both irises. In the more challenging Open Set, where standalone face recognition accuracy dropped to 83.31%, the triple fusion (Face + Left Iris + Right Iris) maintained a high 98.12% accuracy. This demonstrates the "mutual compensation" effect, where the face recognition score can bolster the overall result even when iris quality is degraded.
Conclusion
Fusing NIR face and iris biometrics from a single high-resolution image provides a robust solution for identity verification. The use of NIR lighting mitigates environmental sensitivity, while score-level fusion allows the modalities to mutually compensate for quality deficiencies. Future work will focus on more sophisticated feature-level integration and reducing the computational overhead associated with processing 10MP images in real-time applications.
Summary updated