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

This paper, Detecting Clear Contact Lenses for Iris Recognition: A Two-Stage Mask-Guided Attention Approach by Parisa Farmanifard and Arun Ross, addresses the under-explored problem of detecting clear (transparent) contact lenses in iris recognition. The authors note that while patterned cosmetic lenses are a well-studied presentation attack, clear prescription lenses are typically transparent, make them difficult to detect and are often assumed to have no impact on iris recognition.

The paper first examines this assumption using the commercial VeriEye matcher on four benchmark datasets, showing that clear lenses marginally degrade genuine match scores and increase verification error. Specifically, the authors found that Clear-Clear (CC) genuine pairs score measurably lower than Normal-Normal (NN) pairs on the same subjects, with score gaps ranging from 36 to 39 points across datasets.

To address this, the authors propose a two-stage contact-lens detection framework. The pipeline is formalized as: "Stage 1 uses an existing PAD model to identify patterned lenses, while Stage 2 focuses on the more challenging clear-lens versus no-lens distinction using a ConvNeXt-Base model equipped with Mask-Guided Spatial Attention (MGSA). Stage 1 uses D-NetPAD to classify images as patterned or non-patterned, achieving 100% accuracy with zero bootstrap SD" across all datasets. Only non-patterned images are passed to Stage 2.

Stage 2 introduces the novel MGSA module, which "incorporates a Hough-derived anatomical ROI mask together with learned spatial attention and Squeeze-and-Excitation channel recalibration, allowing the network to focus on subtle limbal cues associated with clear lens wear." The mask is derived from classical circular Hough transform for pupil and limbus localization, creating a three-region binary mask covering the pupil, iris annulus, and an outer arc beyond the limbus. The MGSA output is computed as F ∗ = F ⊙ G ⊙ M̃, where M̃ is the upsampled anatomical mask acting as a hard spatial gate, G is a learned saliency gate from a 1×1 convolution, and the SE block recalibrates channels.

Across the four datasets (UND-AD100, UND-LG4000, IITD-Cogent, IITD-Vista), the full pipeline consisting of both patterned and clear contact lens detection achieves between 90.0%–98.8% accuracy. The authors report that Clear lens accuracy is the lowest of the three classes, confirming that clear lens detection is harder than patterned detection and also harder than normal iris detection. The MGSA module is shown to be critical, as a mask-free baseline degraded by up to 30% on AD100 (60.00% vs. 90.00%). In backbone comparisons, ConvNeXt-Base is the best MGSA backbone, outperforming ResNet-101 and EfficientNet-V2-S, with its 7×7 depthwise kernels better capture the subtle limbal texture evidence of clear lens wear.

Finally, the paper introduces a z-score calibration method that adjusts VeriEye match scores when a clear lens is detected in the input images. The calibration maps scores from each pair type (NN, CC, CN) onto the Normal-Normal reference distribution using the formula scal = (s − µt)/σt · σNN + µNN. This calibration reduces EER by 4.1%–28.3% across datasets, with the largest improvement on IITD-Vista (2.26% → 1.62%). The authors conclude that reliable clear contact lens detection can directly improve iris verification performance, and that lens-aware detection and calibration provide a practical path toward more robust real-world iris recognition.

Improvements to AI Systems:

1. Two-stage hierarchical presentation attack detection (PAD): Instead of a single monolithic classifier, implement a cascaded pipeline where Stage 1 filters out easy-to-detect patterned lenses (using D-NetPAD) and Stage 2 focuses computational resources on the hard clear-lens vs. no-lens discrimination. This improves accuracy and reduces false positives by not forcing one model to handle all attack types.

2. Mask-Guided Spatial Attention (MGSA) module: Integrate a hard anatomical prior (derived from Hough transform for pupil/limbus) as a multiplicative gate alongside learned spatial attention and Squeeze-and-Excitation channel recalibration. This forces the network to attend to limbal regions where clear lens cues reside, improving detection by up to 30% over mask-free baselines.

3. ConvNeXt-Base backbone with large depthwise kernels: Use 7×7 depthwise convolutions instead of standard ResNet or EfficientNet backbones, as they better capture subtle limbal texture variations—critical for transparent lens detection.

4. Z-score calibration for match score correction: When a clear lens is detected, map the raw biometric matcher score from the clear-lens distribution onto the normal-normal reference distribution using the formula scal = (s − µt)/σt · σNN + µNN. This reduces verification EER by 4.1%–28.3% across datasets, directly improving genuine acceptance rates.

5. Score-gap-aware decision thresholds: Use the finding that Clear-Clear genuine pairs score 36–39 points lower than Normal-Normal pairs to dynamically adjust verification thresholds when clear lenses are present, preventing false rejections.

What the Improved AI System Can Do:

• Robust iris recognition in real-world scenarios where users wear transparent prescription lenses, without falsely rejecting them or accepting impostors.

• Automatically classify iris images into three categories (normal, patterned lens, clear lens) with 90–98.8% accuracy, enabling lens-aware processing.

• Self-calibrate biometric match scores on-the-fly based on detected lens type, reducing verification errors without retraining the matcher.

• Operate efficiently by only running the expensive clear-lens detector on non-patterned images, saving compute in high-throughput authentication systems.

• Generalize across different iris sensors (UND, IITD datasets) due to the combination of classical anatomical priors and learned attention, making it deployable in heterogeneous hardware environments.

This work focuses on the impact and detection of clear contact lenses in the context of iris recognition. While the detection of cosmetic or patterned contact lenses has been extensively studied under the presentation attack detection (PAD) paradigm, clear prescription contact lenses, that are typically transparent, have received comparatively less attention despite their widespread use. Unlike patterned lenses, clear lenses introduce no salient texture artifact, making them difficult to detect and are often assumed to have no impact on iris recognition. We first examine this assumption using the commercial VeriEye matcher on four benchmark datasets and show that clear lenses marginally degrade genuine match scores and increase verification error. We then propose a two-stage contact-lens detection framework. Stage 1 uses an existing PAD model to identify patterned lenses, while Stage 2 focuses on the more challenging clear-lens versus no-lens distinction using a ConvNeXt-Base model equipped with Mask-Guided Spatial Attention (MGSA). The proposed MGSA module incorporates a Hough-derived anatomical ROI mask together with learned spatial attention and Squeeze-and-Excitation channel recalibration, allowing the network to focus on subtle limbal cues associated with clear lens wear. Across four datasets, the full pipeline consisting of both patterned and clear contact lens detection achieves between 90.0%--98.8% accuracy. Finally, we introduce a z-score calibration method that adjusts VeriEye match scores when a clear lens is detected in the input images. This calibration reduces EER by 4.1%--28.3% across datasets, demonstrating that reliable clear contact lens detection can directly improve iris verification performance.