EyePreserve: Identity-Preserving Iris Synthesis

Siamul Karim Khan, Patrick Tinsley, Mahsa Mitcheff, Patrick J. Flynn, Kevin W. Bowyer, Adam CzajkaarXiv 2023 · DOI · arXiv

Abstract

Traditional iris recognition systems rely on linear "rubber sheet" models to account for pupil dilation, which often fail to capture complex, non-linear tissue deformations. We present EyePreserve, a fully data-driven framework that synthesizes iris images while strictly preserving identity across a wide range of pupil sizes. By replacing geometric assumptions with a deep learning-based non-linear deformation model, EyePreserve significantly enhances iris recognition performance, particularly in scenarios with high pupil dilation differences.

Introduction

Iris recognition is a cornerstone of biometric identification, yet it faces challenges when comparing images with significantly different pupil sizes—such as those captured in bright versus dark environments. The standard "rubber sheet" model assumes iris tissue stretches linearly. However, real iris muscle fibers exhibit intricate non-linear mechanics during dilation. This oversimplification leads to "identity drift" in synthetic datasets and recognition failures in forensic applications. EyePreserve addresses these challenges by modeling non-linear deformations and generating realistic, identity-consistent iris images.

Methods

The core of EyePreserve is an autoencoder-based deformation model designed to learn iris texture movement directly from data.

The EyePreserve pipeline takes an input iris and a target shape to synthesize a new image with a different pupil size while preserving identity.

The framework employs several key technical innovations:

  • Hybrid Loss Function: To ensure visual realism and biometric consistency, the model integrates perceptual losses (LPIPS, MS-SSIM) with a novel ISO Sharpness loss, ensuring outputs meet international biometric quality standards.
  • Adversarial Identity Preservation: A triplet-based adversarial loss forces the discriminator to learn embeddings where the same identity remains close across different dilations, while distinct identities are separated.
  • Synthesis Integration: The deformation model is paired with StyleGAN architectures (StyleGAN3 for unconditional and StyleGAN2-ADA for conditional synthesis) to create a complete pipeline for generating high-quality iris datasets.

Datasets and Evaluation

The model was trained and validated using diverse datasets, including the Warsaw BioBase Pupil Dynamics (WBPD) for real iris dynamics and CSOSIPAD for identity diversity. External validation on HWS, Q-FIRE, and CASIA-Iris-Lamp datasets confirmed the model's robustness across different sensors and environments.

Results

EyePreserve demonstrates significant improvements over traditional models, particularly as the difference in pupil size (Delta) increases.

Comparison of similarity scores across different models. EyePreserve (gold line) maintains higher similarity than linear or biomechanical models as pupil size difference (Delta) increases.

Key findings include:

  • Superior Recognition: At high dilation (Delta > 0.3), EyePreserve achieved an AUC of 0.996, outperforming linear (0.986) and biomechanical (0.985) models.
  • Identity Stability: Qualitative analysis confirms that EyePreserve maintains specific texture landmarks that standard GANs often lose during pupil size transitions.
  • Robustness: Performance remained consistent across multiple iris matchers (HDBIF and DGR), proving the biometric validity of the learned deformations.
  • Rectification Capabilities: Beyond synthesis, the model can "rectify" diseased irises with irregular pupils back to a standard circular shape while preserving unique identification features.

Conclusion

EyePreserve successfully moves beyond the limitations of linear deformation models by learning complex iris dynamics from data. While it introduces a minute amount of reconstruction noise for very small pupil changes (Delta < 0.1), its performance in high-dilation scenarios is unmatched. Future work will focus on multi-spectral iris synthesis and refining the autoencoding process to eliminate residual noise.

Summary updated

BibTeX
@misc{khan2023eyepreserve,
  author = {Siamul Karim Khan and Patrick Tinsley and Mahsa Mitcheff and Patrick J. Flynn and Kevin W. Bowyer and Adam Czajka},
  title = {{EyePreserve: Identity-Preserving Iris Synthesis}},
  year = {2023},
  eprint = {2312.12028},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.12028},
}
Figure