Ethnicity Prediction Based on Iris Texture Features

Stephen Lagree, Kevin W. BowyerMAICS 2011 · OpenReview

Abstract

This paper presents a robust methodology for predicting ethnicity—specifically Asian versus Caucasian—using iris texture features. By leveraging specialized texture filters and Support Vector Machines (SVM) on normalized iris images, the researchers achieved over 90% classification accuracy. This "soft biometric" approach offers a significant optimization for large-scale identification systems by enabling gallery filtering, thereby reducing search space and computational overhead. Crucially, the study addresses common pitfalls in biometric research, such as data leakage and environmental bias, through person-disjoint validation and standardized imaging hardware.

Introduction

As biometric databases scale to millions of records, the computational cost of matching a probe image against a massive gallery becomes a significant bottleneck. "Soft biometrics"—the automated extraction of secondary traits like gender or ethnicity—can mitigate this by narrowing the search to relevant subsets of the database.

Representative iris image used for texture-based ethnicity classification.

Previous research in iris-based ethnicity prediction often suffered from methodological flaws. "Data leakage," where images of the same individual appear in both training and testing sets, can lead to inflated accuracy by allowing the model to recognize specific individuals rather than general ethnic traits. Additionally, environmental biases, such as varying lighting conditions for different ethnic groups, can cause classifiers to learn camera or lighting artifacts instead of biological features. This study aims to provide a rigorous, bias-free framework for ethnicity prediction.

Methodology

The researchers, Stephen Lagree and Kevin W. Bowyer, developed a pipeline focused on isolating iris texture from confounding factors. The process begins with segmentation and normalization using "IrisBee" software, which transforms the circular iris into a 240x40 pixel rectangular strip. Eyelids and eyelashes are masked to ensure the classifier focuses exclusively on the iris tissue.

Following normalization, nine different texture filters are applied, including spot and line detectors and Laws’ Texture Measures. To capture radial and angular variations, the normalized images are divided into 14 regions. For classification, the Sequential Minimal Optimization (SMO) algorithm was found to be the most effective.

Datasets

The study utilized a controlled dataset from the University of Notre Dame, consisting of 1,200 images from 120 individuals (60 Asian, 60 Caucasian).

Example LG 4000 iris images from subjects with Caucasian ethnicity.

All images were captured using an LG 4000 sensor under near-infrared (NIR) illumination. This standardization ensures that the model learns texture rather than color or sensor-specific noise. To ensure generalizability, the authors employed person-disjoint 10-fold cross-validation, preventing the model from being tested on individuals it encountered during training.

Results and Discussion

The SMO classifier achieved a peak accuracy of 90.58%. In contrast, when person-disjoint constraints were removed, accuracy rose to 96.17%, highlighting how data leakage can lead to overly optimistic results in biometric studies.

Key Findings

  • Algorithm Performance: SMO (90.58%) outperformed Random Forests (89.50%) and Naïve Bayes (68.42%).
  • Texture Dominance: The "E5E5" filter from Laws’ Texture Measures was the most potent single feature, reaching 88% accuracy.
  • NIR Advantage: By using NIR imaging, the system relies on the physical topography of the iris rather than visible color, making it robust across different eye colors.

Conclusion and Future Work

This research demonstrates that iris texture contains reliable ethnic markers that can be automatically extracted with high precision. While the current binary classification is a proof of concept, future work should expand to more diverse ethnic groups and address the complexities of mixed-heritage individuals. Increasing the number of unique subjects in training sets remains a priority for improving the robustness of these soft biometric systems.

Summary updated

BibTeX
@inproceedings{lagree2011ethnicity,
  author = {Stephen Lagree and Kevin W. Bowyer},
  title = {{Ethnicity Prediction Based on Iris Texture Features}},
  booktitle = {MAICS},
  year = {2011},
  url = {https://openreview.net/forum?id=DgSX5cab9kK},
}
Figure