Notre Dame researchers release open-source iris recognition tools built for NIST testing

Researchers at the University of Notre Dame have developed a new open-source toolkit intended to make iris recognition technology more transparent, easier to test, and more accessible to academic researchers working outside the commercial biometric industry.

The paper, Lowering the Barrier to IREX Participation: Open-Source Algorithms, Toolkit, and Benchmarking for Iris Recognition, presents two new iris recognition algorithms, along with open-source implementations designed to comply with the National Institute of Standards and Technology’s (NIST) Iris Exchange, known as IREX.

The work is aimed at a long-standing gap in biometric testing, as NIST’s IREX program has largely evaluated closed-source commercial iris recognition systems rather than open academic tools.

The researchers say that matters because iris recognition is increasingly used in security and identity systems, but many of the most capable algorithms remain proprietary, and that limits outside review, makes reproducibility difficult, and leaves researchers without a strong open baseline for comparing new methods.

It also creates problems for forensic uses of iris recognition, where explainability and human interpretation can be important.

The paper introduces two new neural-network-based methods. The first, called TripletIris, uses a ConvNeXt-tiny model trained with batch-hard triplet loss. In simple terms, the model learns to pull images of the same iris closer together in a mathematical feature space while pushing images of different irises farther apart.

The second, called ArcIris, uses a ResNet100 model trained with ArcFace loss, a method designed to create clearer separation between identities.

The researchers also created IREX-compliant C++ versions of two existing Notre Dame iris recognition methods. One, HDBIF, uses human saliency-driven filtering to encode iris texture. The other, CRYPTS, detects and compares Fuchs’ crypts, visible structures in the iris that can be useful in human-interpretable forensic analysis.

CRYPTS is particularly important from an explainability standpoint, but it also proved too computationally heavy for the strict timing demands of large-scale IREX search.

The toolkit also includes open-source iris segmentation and circle-estimation models. These are the parts of the system that locate the iris and pupil, estimate their boundaries, and prepare the eye image for matching.

That preprocessing step is essential because even a strong recognition algorithm can fail if it cannot properly isolate the iris from eyelids, eyelashes, glare, off-angle images, or other real-world image problems.

A central contribution of the paper is not just that the algorithms exist, but that they were adapted for NIST-style testing. IREX imposes strict requirements on how algorithms handle images, create templates, manage memory, and perform large one-to-many searches.

According to the paper, NIST caps template creation at 1.5 seconds per 640-by-480 image, and a one-to-many search against 500,000 templates returning 50 candidates from both eyes must finish within 25 seconds.

Those constraints shaped the design choices. The researchers used lightweight architectures for segmentation and circle estimation, optimized C++ implementations, and disabled unnecessary threading and gradient tracking to prevent timeouts or memory failures.

ArcIris and TripletIris were able to clear the search-time requirements using fast distance comparisons. HDBIF also met the timing requirements after optimization. CRYPTS, however, failed the timing requirement because it relies on a more expensive comparison process.

The performance results show that modern open-source iris recognition is becoming much more competitive. Across multiple academic datasets, ArcIris and TripletIris substantially outperformed older open-source systems such as OSIRIS and USIT, while in some settings approaching the performance of commercial systems.

ArcIris was generally the stronger and more stable of the two new neural-network methods, especially at strict false-match operating points.

In biometric systems, a low false match rate is critical because it measures how often the system incorrectly says two different people are the same person. But systems must also avoid high false non-match rates, where the system fails to recognize the same person across different images.

The paper found that ArcIris was particularly effective at reducing false non-matches under strict false-match settings.

The researchers also emphasized failure-to-enroll performance. Some algorithms can appear accurate if they simply reject difficult images. But that is less useful in operational environments, where poor image quality, motion blur, occlusion, contact lenses, off-angle capture, and other issues are common.

In the paper’s penalty-based evaluation, systems with high failure rates suffered major performance drops, while ArcIris and TripletIris maintained low failure-to-enroll rates across difficult datasets.

The researchers are not claiming that open-source tools automatically solve the policy and civil liberties concerns surrounding biometric identification. But they are arguing that open systems can make the technology easier to test, reproduce, compare, and scrutinize.

And that is especially important as iris recognition continues to appear in border security, law enforcement, corrections, identity verification, and forensic settings.

The paper’s practical impact may be in lowering the barrier for other researchers to participate in NIST evaluations.

By providing IREX-compliant C++ code, Python implementations, model weights, segmentation tools, and examples of how to satisfy NIST’s technical requirements, the authors created a pathway for academic and open-source teams that may not have had the resources to prepare a formal IREX submission.

The result is a technical paper with policy relevance. It shows that high-performing iris recognition no longer must be confined entirely to closed commercial systems. It also gives researchers, evaluators, and potentially government buyers a clearer way to compare open tools against proprietary systems.

That does not eliminate the need for oversight. Iris recognition remains a powerful biometric technology capable of identifying people at scale. But the paper makes a strong case that if the technology is going to be used, open-source, independently tested, and reproducible systems should be part of the conversation.

Article Topics

biometric dataset  |  biometric testing  |  biometrics  |  biometrics research  |  IREX  |  iris biometrics  |  iris recognition  |  NIST  |  open source