A Comprehensive Evaluation of Iris Segmentation on Benchmarking Datasets.

Area of Science:

• Biometrics and Pattern Recognition
• Computer Vision and Image Analysis
• Artificial Intelligence in Security

Background:

• Iris recognition is a leading biometric modality due to its uniqueness and security.
• Accurate iris segmentation is critical for reliable iris-based biometric authentication.
• Deep learning methods have advanced iris segmentation, but a lack of open-source resources hinders progress.

Purpose of the Study:

• To comprehensively review available open-source iris segmentation resources (datasets, algorithms, tools).
• To benchmark state-of-the-art iris segmentation algorithms against diverse datasets.
• To identify research gaps and foster reproducibility in iris segmentation.

Main Methods:

• Developed and benchmarked three U-Net and U-Net++ inspired segmentation algorithms.
• Trained models on a large composite dataset (>45K samples) with 1K manually segmented ground truth masks.
• Evaluated eleven state-of-the-art algorithms across five datasets with varying conditions.

Main Results:

• Benchmarking revealed strengths and limitations of current iris segmentation algorithms.
• Performance varied significantly across different sensors, environments, and demographic groups.
• Identified key areas for improvement in segmentation accuracy and robustness.

Conclusions:

• The scarcity of open-source resources is a significant challenge in iris segmentation research.
• Publicly releasing developed datasets, algorithms, and tools will facilitate future research and development.
• Standardized benchmarking is essential for advancing the field of iris-based biometrics.