Combining Face and Iris Biometrics for Identity Verification

Yunhong Wang, Tieniu Tan, Anil K. JainAVBPA 2003 · DOI

TL;DR

This paper explores fusing face and iris biometrics to overcome individual system limitations, aiming for improved accuracy and broader user coverage. They propose two fusion strategies: a weighted sum of classifier outputs and treating classifier outputs as a 2D feature vector for a Fisher discriminant or RBF neural network classifier. The fused systems demonstrate superior performance, particularly the RBFNN approach, and can reduce enrollment failure rates.

What Problem Does This Paper Address?

The paper addresses the limitations of standalone biometric systems, specifically face and iris recognition, for identity verification. Face recognition is susceptible to variations in illumination, pose, and expression, while iris recognition can suffer from poor image quality leading to high "failure to enroll" (FTE) rates and user annoyance. Combining these modalities is proposed as a solution to enhance accuracy, increase the population coverage of biometric systems (accommodating individuals with disabilities or those who cannot provide one biometric), and potentially reduce enrollment failures by allowing for simultaneous acquisition and leveraging the strengths of each modality.

Methods

The authors investigate two primary strategies for fusing face and iris classifiers:

  • Weighted Sum Rule: This approach combines the matching distances (outputs) of individual face and iris classifiers. The weights assigned to each classifier are determined by their respective performance metrics (False Accept Rate - FAR, and False Reject Rate - FRR). The idea is to give more weight to classifiers that perform better. The fused score is then compared to a threshold.
  • Feature Vector Approach: Here, the matching distances from the face and iris classifiers are treated as a two-dimensional feature vector (x1, x2). This vector is then fed into a secondary classifier to make the final genuine/impostor decision. Two classifiers are explored for this:
  • Fisher's Discriminant Analysis (FDA): A linear classifier that finds the optimal hyperplane to separate the genuine and impostor classes in the 2D feature space.
  • Radial Basis Function Neural Network (RBFNN): A non-linear classifier that is trained for each individual subject to learn the separation boundary. The paper notes that RBFNNs were chosen due to their superior performance in their experiments.

Datasets

The paper uses two databases for experimental evaluation:

  • Database 1 (DB1): A composite database created by combining publicly available face image datasets (ORL, MIT, Yale) with iris images from the NLPR iris database. To create a paired dataset, arbitrary iris classes were assigned to face images from the public datasets. This database features high-quality iris images.
  • Database 2 (DB2): This database is designed to illustrate the impact of enrollment failure. It includes some poor-quality iris images that would typically be rejected by an operational system, alongside the ORL face database. This allows for the evaluation of how fusion handles imperfect biometric data.

Key Insights and Ideas

  • Fusion Enhances Robustness: Combining face and iris biometrics significantly improves the overall verification accuracy and robustness compared to individual systems, especially when dealing with imperfect or low-quality data.
  • Broader User Coverage: Fusion is not just about accuracy; it's crucial for increasing the number of individuals who can successfully use a biometric system, accommodating diverse needs and abilities.
  • Reduced Enrollment Failure: The ability to acquire both face and iris images simultaneously, and the robustness of fused systems to occasional poor-quality iris images, can lead to a lower effective enrollment failure rate.
  • RBFNN for Subject-Specific Fusion: Training individual RBFNN classifiers for each subject proved to be the most effective fusion strategy, achieving the highest verification accuracy. This suggests that personalized fusion models can capture subtle differences between genuine and impostor scores for each individual.
  • Weighted Sum Rule Offers a Good Trade-off: While RBFNN achieved the highest accuracy, the weighted sum rule also provided substantial improvements and a wider range of operating points with zero errors, making it a practical and effective fusion method.

Key Results

  • DB1 Performance: The fused systems consistently outperformed individual face and iris classifiers. The weighted sum rule and Fisher's discriminant analysis showed improved separation of genuine and impostor scores. The RBFNN-based fusion achieved the highest verification accuracy.
  • DB2 Performance: On DB2, which included poor-quality iris images, the fusion systems were still able to achieve high accuracy. Notably, the RBFNN fused system correctly identified images that the standalone iris classifier failed on, demonstrating the benefit of fusion in handling data that might otherwise lead to enrollment failure.
  • Error Rates: Table 1 shows that for DB1, the fused systems (especially weighted sum and Fisher) achieved significantly lower total error rates (FAR+FRR) at various thresholds compared to individual systems. The RBFNN approach achieved zero error rates across a wide range of thresholds.

Limitations and Future Work

  • Limitations: The paper acknowledges that the performance on DB2 (with poor quality iris images) was not as high as on DB1. While fusion helps, it doesn't completely eliminate the impact of poor data. The RBFNN approach, while highly accurate, requires training a separate network for each subject, which can be computationally intensive for large-scale systems.
  • Future Work: The paper implicitly suggests that further research into more efficient RBFNN training or alternative subject-specific fusion methods could be beneficial. The successful demonstration of fusion's ability to mitigate enrollment failure and improve accuracy points towards its continued importance in developing robust and inclusive biometric systems.

Summary updated

BibTeX
@inproceedings{wang2003combining,
  author = {Yunhong Wang and Tieniu Tan and Anil K. Jain},
  title = {{Combining Face and Iris Biometrics for Identity Verification}},
  booktitle = {AVBPA},
  year = {2003},
  doi = {10.1007/3-540-44887-x\_93},
}
Figure