Abstract
Intra-class variations—changes in a person's appearance due to factors like aging, lighting, or posture—significantly degrade the performance of biometric systems over time. While the impact of these variations is well-documented, effective solutions for template updating remain scarce. Most existing methods are either trait-specific or vulnerable to "template poisoning" by outliers. This paper proposes a novel, generic template update framework that utilizes "pass tables" to accumulate information for outlier removal and representative template selection. By applying a novel criterion to select the optimal template set based on similarity thresholds, the framework ensures high accuracy and robustness. Experimental results across synthetic and real datasets for various biometric traits demonstrate that this method outperforms existing techniques, achieving lower False Rejection Rates (FRR) and maintaining system integrity against outliers.
Introduction: The Challenge of Intra-Class Variation
Biometric recognition systems, such as face or fingerprint scanners, rely on comparing a live sample against a stored "template." However, human appearance is not static. Factors such as aging, changes in posture, and varying environmental lighting conditions introduce intra-class variations. These variations cause the initial stored templates to become less representative over time, leading to a steady increase in the False Rejection Rate (FRR).
To combat this, systems must periodically update their template sets. The challenge lies in selecting new templates that capture genuine variations without accidentally incorporating "outliers" or imposters—a phenomenon known as "template poisoning" or "creep-in errors." Furthermore, most current update strategies are designed for specific biometric traits (e.g., only fingerprints), lacking the flexibility to be applied across diverse biometric modalities.
The Generic Template Update Framework
The proposed framework addresses these issues through a robust, three-pillar approach designed to be "plug-and-play" for any biometric system using nearest-neighbor classifiers.
1. Candidate Selection and Closeness Thresholds
Instead of simply accepting any sample that passes a basic matching threshold, the system introduces a "closeness threshold." If a new sample is too similar to existing templates, it is discarded as redundant. If it is sufficiently different but still valid, it is moved to a candidate template set ($T_c$) for further evaluation. This ensures the template set remains lean and computationally efficient.
2. The Innovation of Pass Tables
The core of the framework is the Pass Table. This is a binary matrix that records whether specific samples "pass" or "fail" against various templates. By tracking these interactions over time, the system can identify which templates are most effective at recognizing the user's variations and which are redundant or potentially erroneous.
3. Robust Selection Algorithm
The framework merges sample and template pass tables into a "Joint Pass Table." An optimization algorithm then selects a set of $K$ templates that provides maximum "permeability." This mathematical approach identifies the set that covers the widest possible range of genuine user variations while effectively filtering out outliers that do not align with the broader data pattern.
Experimental Validation
The researchers tested the framework across several diverse datasets to prove its versatility and effectiveness:
- Synthetic 2D Points: Used to visualize clustering and outlier rejection compared to classical methods like k-Means.
- Synthetic Palmprint Dataset: 571 images with simulated noise and variations over 31 sessions.
- Real Palmprint & Palmvein Datasets: A long-term study of 60 subjects tracked over 12 months to test resistance to "template aging."
- CMU Multi-PIE Face Database: A benchmark containing 337 subjects under 15 viewpoints and 19 lighting conditions.
Key Results and Performance
The framework demonstrated significant improvements in both accuracy and security across all tested modalities:
- Face Recognition Gains: On the Multi-PIE dataset, the framework successfully updated templates to handle extreme lighting and pose changes, reducing the FRR from an initial 33.73% to just 14.2% after several update cycles.
- Superior Accuracy: In synthetic palmprint trials, the proposed method achieved an FRR of 21.28%, vastly outperforming the "No-update" baseline of 61.52% and exceeding the performance of existing methods like DENT (23.51%).
- Zero Creep-in Errors: Unlike k-Means or DENT, which occasionally accepted outliers as valid templates, the proposed framework maintained 0% creep-in errors, ensuring system security was never compromised during the update process.
Conclusion and Future Outlook
This research provides a robust, mathematically grounded solution to the persistent problem of template aging and intra-class variation. By shifting the focus from trait-specific features to a logical coverage problem via Pass Tables, the authors have created a universal module for biometric maintenance.
While the current selection algorithm is highly effective, future work may focus on optimizing its computational complexity for extremely large template sets. Additionally, the researchers aim to expand testing to other biometric modalities, such as iris and gait recognition, to further cement the framework's status as a truly generic solution for the biometric industry.
Summary updated