Abstract
Researchers at the University of Notre Dame have developed the first automated classifier designed to determine if two iris images belong to monozygotic (identical) twins. By employing a Siamese network and contrastive learning, the study demonstrates that deep learning can achieve approximately 81% accuracy in identifying these subtle genetic links. Notably, the model finds more useful genetic signals in the surrounding ocular structure—such as eyelids and skin—than in the iris texture itself, which is traditionally considered a random biometric.
Introduction
In biometric identification, "Daugman-style" iris recognition is the gold standard, predicated on the fact that iris textures are formed randomly during fetal development. Consequently, even the left and right eyes of the same individual are as distinct to a computer as those of a stranger. However, previous research indicated that humans can distinguish monozygotic (MZ) pairs from non-monozygotic (NMZ) pairs with roughly 80% accuracy, suggesting a "genetic signature" hidden in the eye region that traditional algorithms overlook. This study aims to automate zygosity detection, a task that otherwise requires expensive DNA testing.
Methods
The researchers utilized a Siamese Network architecture, which consists of two identical neural networks sharing the same weights to process a pair of images simultaneously.
- Backbone: A modified ResNet-18 serves as the encoder, producing a 128-dimensional embedding for each image.
- Learning Strategy: The model is trained using contrastive learning, minimizing the Euclidean distance between embeddings of MZ pairs while maximizing the distance for NMZ pairs.
- Ablation Study: To isolate the source of the genetic signal, three versions of the model were trained: raw images (full eye region), iris-only (masked surrounding), and non-iris-only (masked iris).
Datasets
The study leveraged two primary datasets from the University of Notre Dame:
- ND-IRIS-0405: Used to create "synthetic" MZ pairs (left and right eyes of the same person) for training, providing over 23,000 image pairs.
- BXGrid: A repository of images from actual identical twins ("natural" MZ pairs), used for the final test set to ensure real-world generalization.
- Quality Control: Images were filtered using ISO/IEC 29794-6 metrics to maintain high standards for lighting and focus.
Key Insights and Results
The performance evaluation across the three input types revealed a surprising "periocular surprise." The model trained on the surrounding eye region (non-iris-only) outperformed the iris-only model, achieving 77% accuracy versus 73%. This indicates that while iris texture is highly random, the surrounding ocular structure—eyelid shape, skin folds, and sclera patterns—contains stronger genetic markers.
| Model Input | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|
| Original (Raw) | 0.81 | 0.82 | 0.81 | 0.81 |
| Non-iris-only | 0.77 | 0.84 | 0.68 | 0.74 |
| Iris-only | 0.73 | 0.80 | 0.60 | 0.69 |
The full-image model achieved 81% accuracy, slightly exceeding the reported 80% accuracy of human experts. A significant challenge identified was "dilation noise"; performance drops when pupil dilation levels differ between the two images.
Discussion and Future Work
Future work will focus on expanding the "natural" twin test set with more diverse demographic groups and implementing dilation normalization to standardize iris images. Additionally, the authors suggest using explainability tools like Grad-CAM to pinpoint the exact periocular features the model relies on for its decisions.
Conclusion
This research demonstrates that deep learning can effectively detect genetic relationships through ocular imagery, revealing that the eye region contains more than just random biometric data. By fusing information from both the iris and the periocular area, the Siamese network matches human-level performance in zygosity detection.
Summary updated