Abstract
Multibiometric fusion has been an active research area for many years. Score normalization transforms scores from different matchers into a common domain. This paper surveys classical score normalization techniques and recent advances. The performance of normalization functions—MinMax, Tanh, Zscore, PL, LTL, RHE, and FF—is evaluated using the XM2VTS Benchmark. Performance is measured via EER, AUC, GAR (at FAR=0.001), and EER threshold stability. Results indicate that no single technique performs best for all systems, though PL and FF normalization outperform others in many applications.
Introduction
In biometric security, relying on a single modality (e.g., fingerprint or face) can be unreliable due to noise or spoofing. Multibiometric systems mitigate this by fusing information from multiple sources, most commonly at the score level. However, matchers are heterogeneous; one might output a probability [0, 1] while another outputs a distance [0, 1000]. To combine these meaningfully, they must be transformed into a common domain through score normalization. This process must be robust to outliers while preserving the discriminative power of the original data.
Normalization Techniques
The study compares three classical and four advanced methods:
- Classical Methods: MinMax (scaling to [0, 1]), Z-score (mean and standard deviation based), and Tanh (a robust sigmoid-like mapping).
- Advanced Methods: Piecewise Linear (PL), Linear Tanh Linear (LTL), Reduction of High-scores Effect (RHE), and FRR/FAR-based (FF) normalization.
Following normalization, the Sum Rule (adding the normalized scores) is applied for final fusion and decision-making.
Experimental Setup
Evaluation was conducted on the XM2VTS-Benchmark database, featuring five face matchers and three speech matchers. Following the Lausanne Protocol-1 (LPI), the data was split into training and evaluation sets. Fifteen distinct multimodal combinations were tested to ensure a comprehensive evaluation of normalization performance across varying biometric distributions.
Results and Discussion
The performance of these methods varies significantly across different fusion scenarios and metrics. Figure 3 illustrates the variability in Equal Error Rate (EER) across the 15 tested combinations, highlighting that no single method consistently dominates.
To aggregate these results, a "Performance Mark" system was used, ranking methods from 1 to 7 for each test. As shown in Figure 6, the optimal method depends on the specific performance goal.
- Error Minimization (EER): The FF (FRR/FAR based) method was the clear leader with a total mark of 81.
- Discriminative Power (AUC): The PL (Piecewise Linear) method performed best with a mark of 82.
- High Security (GAR at FAR=0.001): PL again led the pack with 78.5.
- Stability: FF, LTL, and Tanh demonstrated the most stable thresholds (TE), which is vital for real-world deployment where environmental conditions vary.
Conclusion
There is no "one-size-fits-all" normalization technique. While PL and FF are the most versatile and high-performing methods for most modern multibiometric applications, the choice of method should be driven by system priorities—whether minimizing error rates or maximizing overall accuracy. Future work should explore these techniques under different fusion rules (e.g., Product, Max) and investigate why simpler models like PL occasionally outperform more complex successors like LTL.
Summary updated