DeformIrisNet: An Identity-Preserving Model of Iris Texture Deformation

Siamul Karim Khan, Patrick Tinsley, Adam CzajkaWACV 2023

TL;DR

This paper introduces DeformIrisNet, a novel deep learning model that learns to deform iris textures to match varying pupil sizes while preserving identity. This addresses a key challenge in iris recognition where pupil dilation causes significant within-class variance, leading to improved accuracy and potential applications in forensic analysis.

What Problem Does This Paper Address?

Iris recognition is a highly reliable biometric modality due to the unique and stable patterns of the iris. However, a major source of error in iris recognition systems is the variation in pupil size. As the pupil dilates or constricts, the iris texture is non-linearly deformed. Traditional iris recognition methods often normalize the iris by mapping it to a rectangular region, but this linear approach fails to accurately compensate for large pupil size variations. While some non-linear methods exist, they often rely on biomechanical assumptions or are simplifications of the complex iris muscle movements. This paper tackles the problem of accurately modeling and correcting these complex, non-linear iris texture deformations directly from data, without relying on prior assumptions about iris biomechanics or precise pupil size definitions.

Methods

The core of this work is DeformIrisNet, a deep autoencoder architecture based on a modified U-Net. The model takes an input iris image and a target iris mask (defining the desired shape) and generates a new iris image that matches the target shape. The novelty lies in its training mechanism, which employs a multi-component loss function:

  • Identity-Preserving Loss: This component uses domain-specific human-inspired filters (Firis) to extract features from both the generated and target iris images. By minimizing the difference in these features, the model is encouraged to preserve the identity of the iris.
  • Mask Loss: This loss uses a lightweight segmentation network (CC-Net) to ensure the generated iris matches the shape specified by the target mask.
  • Perceptual Loss (LPIPS): This loss leverages deep features from AlexNet to ensure the generated image is perceptually similar to the ground truth.
  • L1 Loss: This component directly minimizes the pixel-wise difference between the generated and target images, aiming for visual realism.

The architecture itself features modifications to the standard U-Net, including replacing max pooling with strided convolutions and bilinear downsampling for better feature capture, and using sub-pixel convolution for upsampling to reduce checkerboard artifacts.

Datasets

The primary dataset used is the Warsaw-BioBase-Pupil-Dynamics v3.0 (WBPD) dataset. This dataset consists of near-infrared iris videos captured under varying visible light conditions, leading to natural pupil dilation and constriction. The authors curated this dataset by pairing images with different pupil-to-iris ratios to create input-target pairs for training. The data was split into eye-disjoint training, validation, and test sets to ensure the model learns generic iris muscle movements rather than subject-specific dynamics.

Key Insights and Ideas

  • Data-Driven Deformation: The model learns complex iris deformations directly from data, bypassing the need for biomechanical models or explicit pupil size measurements, which are often difficult to define accurately.
  • Identity Preservation is Key: The inclusion of an identity-preserving loss component is crucial. Without it, generative models can produce visually realistic but identity-altering deformations, rendering them useless for biometrics.
  • U-Net Architecture Suitability: While initial experiments with StyleGAN3 were disappointing, the U-Net architecture, with specific modifications for downsampling and upsampling, proved effective for this task.
  • "By-product" Capabilities: The model exhibits emergent capabilities in iris inpainting (filling in occluded regions, like eyelids) and noise reduction, suggesting its potential beyond core deformation correction.

Key Results

  • Improved Iris Recognition: When used to rectify iris images before recognition, DeformIrisNet achieved a higher decidability score (d') of 3.003 compared to linear (2.642) and biomechanical (2.644) methods, indicating better separation between genuine and imposter scores.
  • Comparable Equal Error Rate (EER): While the EER was similar across methods (around 0.115-0.118), the ROC curves showed that DeformIrisNet achieved a better recognition accuracy (higher Area Under the ROC Curve).
  • Visual Realism: The generated iris images are visually realistic and preserve identity, making them suitable for both automated recognition and human forensic examination.

Limitations and Future Work

  • Limited Application in Forensics: While the model can generate iris images with matching pupil sizes for human examination, the generated texture for regions not visible in the original image is "dreamed up" and not guaranteed to represent the true identity. This limits its application to cases where a significant portion of the iris is visible.
  • Quantitative Human Evaluation: The paper suggests that human examination experiments are needed to quantitatively assess the benefits of the proposed normalization for human examiners.
  • "Denoising" Task: Training specifically for iris denoising could yield more interesting results, as the current model's denoising capability is an emergent property.
  • Potential for "Painting" Iris Images: The model's ability to generate realistic iris images for arbitrary masks suggests potential for "painting" iris images, though this was not the primary optimization goal.

Summary updated

BibTeX
@inproceedings{khan2023deformirisnet,
  author = {Siamul Karim Khan and Patrick Tinsley and Adam Czajka},
  title = {{DeformIrisNet: An Identity-Preserving Model of Iris Texture Deformation}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2023},
}
Figure