[Paper Review] Gender-From-Iris or Gender-From-Mascara?

Andrey Kuehlkamp, Benedict Becker+1 more|arXiv (Cornell University)|Feb 4, 2017

Face recognition and analysis14 references4 citations

TL;DR

This paper investigates gender classification from iris images using multi-layer perceptrons (MLP) and convolutional neural networks (CNN), comparing data-driven and hand-crafted features under person-disjoint training/testing with and without makeup. It finds that gender-from-iris accuracy is significantly lower than previously reported—around 66% with LBP features—due to confounding factors like cosmetics and data leakage from non-person-disjoint splits.

ABSTRACT

Predicting a person's gender based on the iris texture has been explored by several researchers. This paper considers several dimensions of experimental work on this problem, including person-disjoint train and test, and the effect of cosmetics on eyelash occlusion and imperfect segmentation. We also consider the use of multi-layer perceptron and convolutional neural networks as classifiers, comparing the use of data-driven and hand-crafted features. Our results suggest that the gender-from-iris problem is more difficult than has so far been appreciated. Estimating accuracy using a mean of N person-disjoint train and test partitions, and considering the effect of makeup - a combination of experimental conditions not present in any previous work - we find a much weaker ability to predict gender-from-iris texture than has been suggested in previous work.

Motivation & Objective

• To re-evaluate gender-from-iris accuracy under realistic experimental conditions, including person-disjoint training/testing and makeup variation.
• To assess whether iris texture truly contains gender-discriminative information or if performance is driven by confounding factors like cosmetics or image intensity.
• To compare the performance of data-driven features (learned by MLPs and CNNs) versus hand-crafted features (LBP, Gabor, intensity) in gender classification.
• To investigate whether CNNs offer performance gains over MLPs and hand-crafted features in gender-from-iris tasks.
• To quantify the impact of occlusion masks and image intensity thresholds on gender prediction accuracy.

Proposed method

• Used a person-disjoint 80/20 train-test split across 2500 batches of 32 images, with 20 randomized trials per condition to reduce variance.
• Employed multi-layer perceptrons (MLP) and convolutional neural networks (CNN) with ReLU activation, training on full eye images and segmented irises at three resolutions.
• Extracted data-driven features via end-to-end learning in MLP and CNN architectures, and hand-crafted features using intensity, Gabor filtering, and Local Binary Patterns (LBP).
• Evaluated performance on full eye images and segmented irises, including occlusion masks and thresholded intensity maps.
• Compared results across different input resolutions (e.g., 95×120 to 285×360) and eye types (left/right) to assess sensitivity to image scale and modality.
• Used 10-fold cross-validation and mean accuracy across trials to ensure robustness and reduce overfitting bias.

Experimental results

Research questions

• RQ1 How does person-disjoint training and testing affect the estimated accuracy of gender-from-iris systems compared to non-person-disjoint splits?
• RQ2 To what extent does eye makeup bias gender classification accuracy in iris-based systems?
• RQ3 Can data-driven features from MLPs and CNNs outperform hand-crafted features like LBP or Gabor filtering in gender-from-iris tasks?
• RQ4 Does the use of segmented irises or occlusion masks contribute significantly to gender prediction, independent of actual iris texture?
• RQ5 What is the true gender classification accuracy when isolating iris texture from confounding factors like image intensity and cosmetics?

Key findings

• Using person-disjoint training and testing, the mean gender classification accuracy for hand-crafted LBP features with MLP was 66% ± 2.7%, significantly lower than previously reported accuracies.
• Data-driven features from MLPs and CNNs yielded only ~60% accuracy, indicating that learned features do not improve performance over simple intensity or thresholding methods.
• Thresholding the average iris intensity achieved approximately 60% gender classification accuracy, suggesting that low-level image statistics dominate over texture details.
• Occlusion masks alone (excluding iris texture) achieved ~60% accuracy, indicating that gender prediction is heavily influenced by non-textural features like eyelid shape or eyelash patterns.
• CNNs showed no significant performance gain over MLPs, with similar accuracy across all resolutions and input types, implying that the network does not extract higher-level texture features.
• The presence of eye makeup increased estimated accuracy, indicating that cosmetics introduce a confounding factor that inflates performance in non-person-disjoint evaluations.