🤖 AI

Detecting Clear Contact Lenses for Iris Recognition: A Two-Stage Mask-Guided Attention Approach

This paper addresses the overlooked challenge of detecting transparent clear contact lenses in iris recognition by demonstrating their negative impact on verification accuracy and proposing a two-stage framework that combines patterned lens detection with a novel Mask-Guided Spatial Attention model to identify clear lenses, ultimately improving system performance through score calibration.

Original authors: Parisa Farmanifard, Arun Ross

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Parisa Farmanifard, Arun Ross

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your eye is a unique fingerprint, but instead of ridges on your skin, it's a swirling, intricate map of colors and lines on the surface of your eye. This map is called the iris, and it's so detailed that it can tell computers who you are, even distinguishing between identical twins. For years, scientists have trusted this map to be the ultimate ID card. But there's a catch: what if someone puts on a disguise? Just like a person might wear a mask to hide their face, someone could wear a contact lens to trick the eye scanner. While we know about "cosmetic" lenses that paint wild patterns over the eye, there's a sneaky, invisible kind of disguise: clear contact lenses. These are the everyday lenses people wear to see better, which are transparent and leave no obvious pattern. For a long time, experts assumed these clear lenses were harmless to the scanner, like wearing clear glasses. But this paper asks a simple, nagging question: Is that really true, or is the scanner getting fooled by something it can't see?

This research, presented by Parisa Farmanifard and Arun Ross from Michigan State University, dives deep into the world of "Presentation Attack Detection" (PAD), which is just a fancy way of saying "spotting the fake." The authors challenge the old assumption that clear contact lenses don't matter. They set out to prove two things: first, that these invisible lenses actually mess up the scanner's ability to recognize you, and second, that they can build a super-smart detective system to find them.

The Invisible Disguise

The team started by testing a commercial eye-scanning system (called VeriEye) on four different sets of data. They wanted to see if clear lenses were actually hurting the system's performance. The results were a bit of a shocker. Even though the lenses are clear, they do introduce tiny, almost invisible distortions at the edge of the iris (where the colored part meets the white part). These tiny glitches act like static on a radio signal. The study found that when someone wears a clear lens, the scanner's "genuine match score" (how sure it is that it's you) drops slightly. It's not a total failure, but it's enough to make the system more likely to say "No, that's not you" when it actually is. In short, the invisible lens does leave a trace, and it makes the scanner less confident.

The Two-Stage Detective

Since clear lenses are so hard to spot, the authors decided that a single "look and guess" approach wouldn't work. Instead, they built a two-stage detective team, like a security checkpoint with two different guards.

Stage 1: The Pattern Hunter
The first guard is a specialist in spotting the obvious fakes. This stage looks at the eye to see if there are any bold, colorful patterns (like the kind used in Halloween costumes or fashion lenses). These are easy to catch because they look like a painted texture over the eye. The team used an existing, highly trained AI model for this job. If the eye has a pattern, the system flags it immediately. This stage is incredibly good at its job, catching 100% of the patterned lenses in their tests.

Stage 2: The Micro-Scope
If the first guard says, "No patterns here, it looks normal," the eye gets passed to the second, more difficult guard. This guard's job is to find the clear lenses. This is the hard part because clear lenses don't have patterns; they only have a tiny, faint reflection at the very edge of the iris. To find this, the team invented a new trick called Mask-Guided Spatial Attention (MGSA).

Think of the eye image as a giant, noisy room. A normal AI might get distracted by the furniture in the background (the eyelids, the skin around the eye). The MGSA trick is like handing the AI a pair of glasses with a specific "zone" highlighted on the lens. This zone is a map of exactly where the iris and its edge should be. The AI is forced to ignore everything outside this zone and focus its super-powerful attention only on that tiny ring where the clear lens might be hiding. It's like telling a detective, "Don't look at the whole house; just check the front door for a scratch."

The team tested this new detective system on four different datasets (collections of eye images) and found it worked remarkably well. The full two-stage system correctly identified the type of lens (Patterned, Clear, or Normal) between 90.0% and 98.8% of the time. They also tested different "brains" for the AI and found that a model called ConvNeXt-Base was the best at spotting those tiny edge clues, outperforming other popular models.

Fixing the Score

The paper doesn't just stop at finding the lenses; it also fixes the problem they cause. Since the authors proved that clear lenses lower the scanner's confidence score, they created a "calibration" method. Imagine the scanner gives you a score of 850, but because you're wearing a clear lens, it drops to 810. The system now knows to say, "Ah, this person is wearing a clear lens, so I should bump that score back up to 850 to be fair."

By using their two-stage detector to guess if a lens is present and then adjusting the score accordingly, they were able to significantly improve the system's accuracy. In the best case (on the IITD-Vista dataset), this simple fix reduced the error rate by 28.3%. This means the system becomes much better at letting real users in while keeping impostors out, simply because it learned to account for the invisible disguise.

The Bottom Line

This paper shows that clear contact lenses are not as invisible to biometric scanners as we thought. They do cause a measurable drop in performance, but by using a smart, two-step detection system that focuses on the specific edge of the eye, we can find them and correct the scanner's mistakes. The authors suggest that by adding this "lens-aware" step, we can make iris recognition much more reliable for everyone, whether they are wearing glasses, contacts, or nothing at all. It's a reminder that in the world of security, even the most transparent disguises leave a trace if you know exactly where to look.

Technical Summary: Detecting Clear Contact Lenses for Iris Recognition

Problem Statement
While the detection of patterned (cosmetic) contact lenses in iris recognition is well-established under the Presentation Attack Detection (PAD) paradigm, clear prescription contact lenses remain a significant, underexplored challenge. Unlike patterned lenses, which introduce salient synthetic textures, clear lenses are optically transparent and leave no obvious visual artifacts, leading to the assumption that they do not impact iris recognition. However, the authors challenge this assumption, noting that clear lenses introduce subtle refractive index mismatches and edge artifacts at the iris-sclera limbus. These subtle cues degrade genuine match scores and increase verification errors, yet they are difficult to detect due to optical invisibility, extreme spatial localization (confined to a narrow annular band), and high intra-class variability across lens brands and acquisition conditions.

Methodology
The authors propose a two-stage detection framework designed to address the fundamental disparity between detecting patterned lenses (a texture-based problem) and clear lenses (a fine-grained spatial localization problem).

Stage 1: Patterned Lens Screening
- Input: A compact, localized iris crop ( $\hat{I}$ ).
- Model: An existing PAD model, D-NetPAD (based on DenseNet-121), is fine-tuned to classify images as "patterned" or "non-patterned."
- Logic: Images identified as patterned are immediately classified. Only "non-patterned" images proceed to Stage 2, ensuring the second stage focuses exclusively on the more difficult clear-vs-normal distinction without being dominated by easy texture cues.
Stage 2: Clear Lens Detection via Mask-Guided Attention
- Input: The full periocular image ( $I$ ) and an automatically generated anatomical ROI mask ( $M$ ).
- Mask Generation: A binary mask is derived using classical circular Hough transforms to localize the pupil and limbus. The mask defines three regions: the pupil, the iris annulus, and an "outer arc" (the periocular tissue immediately beyond the limbus) where clear lens evidence is most likely to appear. No pixel-level annotations or learned segmentation are used.
- Architecture: The core of Stage 2 is a Mask-Guided Spatial Attention (MGSA) module integrated into a ConvNeXt-Base backbone.
  - Hard Anatomical Gate: The upsampled mask $\tilde{M}$ is multiplied element-wise with feature maps to zero out activations outside the anatomically relevant regions (iris and outer arc), preventing the network from attending to irrelevant background features.
  - Learned Saliency Gate: A $1\times1$ convolution generates a soft spatial gate $G$ to re-weight specific areas within the masked region, learning to focus on the subtle limbal cues.
  - Channel Recalibration: A Squeeze-and-Excitation (SE) block globally recalibrates feature channels to select the most discriminative features.
- Backbone Comparison: The authors evaluated EfficientNet-B4, ResNet-101, and ConvNeXt-Base. ConvNeXt-Base was selected as the optimal backbone, attributed to its large-kernel ( $7\times7$ ) depthwise convolutions which better capture the spatially extended, subtle curvature of the limbal boundary.
Score Calibration
- Upon detecting a clear lens, the system applies a z-score calibration to the raw VeriEye verification scores.
- Using the predicted lens labels (Normal, Clear), the system estimates the mean and standard deviation of genuine pair scores for each condition.
- Scores are rescaled to align with the "Normal-Normal" reference distribution, effectively closing the score gap introduced by the lens without altering the rank ordering within pair types.

Key Contributions

Quantitative Evidence of Degradation: The study provides empirical evidence across four large-scale datasets that clear lenses systematically degrade genuine match scores in commercial matchers (VeriEye), increasing verification error rates.
Two-Stage Detection Pipeline: The authors decompose the three-class problem into two specialized binary decisions. This approach significantly outperforms single-stage baselines by isolating the texture anomaly of patterned lenses from the spatial localization challenge of clear lenses.
Novel Mask-Guided Attention (MGSA): The paper introduces an annotation-free method that injects a Hough-derived anatomical ROI mask directly into the deep backbone's feature hierarchy. This combines a fixed spatial prior with learned attention, forcing the network to focus on the specific limbal regions where clear lens evidence exists.
Verification Improvement: The study demonstrates that integrating lens detection with score calibration yields substantial improvements in verification performance, reducing Equal Error Rates (EER) by 4.1% to 28.3% across datasets.

Results

Classification Accuracy: The full pipeline achieved accuracy between 90.0% and 98.8% across four datasets (UND-AD100, UND-LG4000, IITD-Cogent, IITD-Vista).
- Stage 1 achieved 100% accuracy in detecting patterned lenses across all datasets.
- Stage 2 (Clear vs. Normal) achieved accuracies ranging from 87.5% (UND-AD100) to 98.18% (IITD-Vista).
- ConvNeXt-Base with MGSA consistently outperformed ResNet-101 and EfficientNet baselines, particularly on larger datasets.
Verification Performance: The proposed z-score calibration reduced the EER significantly:
- UND-AD100: 0.49% $\to$ 0.47% (4.1% relative reduction).
- IITD-Vista: 2.26% $\to$ 1.62% (28.3% relative reduction).
- The improvement correlated with the severity of the initial score degradation; datasets with larger score gaps saw greater benefits from calibration.
Generalization: The method showed robust cross-sensor generalization, particularly when models trained on larger datasets (IITD) were tested on others. However, models trained on smaller datasets (UND) struggled to generalize to IITD sensors, highlighting an asymmetric transfer gap.

Significance and Claims
The paper claims that reliable detection of clear contact lenses is not merely a security measure against spoofing but a critical component for improving the robustness of real-world iris recognition. By demonstrating that clear lenses introduce a systematic, condition-dependent bias rather than random noise, the authors argue that lens-aware detection followed by score calibration is a practical and effective path to mitigating verification errors. The work establishes that anatomical priors (via Hough masks) can effectively guide deep networks to detect subtle, invisible artifacts, offering a solution to a problem that has remained open for over a decade. The authors note that while the current system relies on Hough-based masks and near-infrared data, the framework provides a foundation for more robust biometric systems in diverse operational environments.

The Invisible Disguise

The Two-Stage Detective

Fixing the Score

The Bottom Line

More like this