Two-headed eye-segmentation approach for biometric identification

Wiktor Lazarski, Maciej Zieba, Tanguy Jeanneau, Tobias Zillig, Christian BrendelarXiv 2022 · OpenReview

TL;DR

This paper introduces a novel "two-headed" deep learning architecture for segmenting key eye components (iris, pupil, eyeball) and eyelashes, crucial for robust iris-based biometric identification. By separating these tasks into distinct decoding heads, the model can effectively incorporate a convex prior for more accurate iris segmentation and explore various loss functions for improved performance.

What Problem Does This Paper Address?

Iris-based biometric identification is a highly accurate method for person identification, but it relies heavily on precise segmentation of the eye region. Existing segmentation methods often struggle with accurately identifying all necessary components, particularly the eyelashes, which can be noisy or inconsistently annotated. This paper addresses the challenge of achieving high-quality, robust segmentation of both the core eye structures and the surrounding eyelashes, which is essential for reliable iris recognition systems. The authors highlight that previous approaches often lack comprehensive ablation studies on loss functions and the impact of incorporating priors like convexity.

Methods

The core innovation is a two-headed architecture built upon the standard DeepLabV3 framework. While sharing a common encoder for feature extraction, the model splits into two distinct decoding heads:

  • Geometry Head: Responsible for segmenting the main eye components: eyeball, iris, and pupil.
  • Noise Head: Specifically designed to segment eyelashes, treated as "noise" components that can interfere with identification.

This separation allows for specialized processing and the application of different techniques. A key aspect is the investigation of various loss functions for training, including:

  • Dice Loss: A common metric for segmentation that measures overlap.
  • Boundary Loss: Focuses on correctly identifying the edges of segmented regions.
  • Surface Loss: Utilizes signed distance functions to penalize errors based on distance from the true boundary.

Furthermore, the paper explores the use of a convex prior for the geometry head. This prior encourages the segmented iris and pupil shapes to be convex, a property they naturally possess. The authors implement this as a differentiable component that can be trained end-to-end, either as a post-processing step or integrated directly into the training.

Datasets

The study utilizes a dataset comprising 11,560 training, 1,448 validation, and 1,420 test high-resolution near-infrared eye images. Each image is annotated with geometry segmentation masks (pupil, iris, sclera) by three annotators, and a separate mask for noise regions (eyelashes). The final ground truth masks are derived from majority voting. The paper also discusses a preprocessing step for eyelashes to standardize their annotation width, addressing inconsistencies from annotators.

Key Insights and Ideas

  • Decoupling Segmentation Tasks is Beneficial: Separating the segmentation of core eye components from eyelashes into distinct heads allows for more specialized and effective learning for each task.
  • Convexity Prior for Iris Accuracy: Enforcing convexity on the iris and pupil segmentation significantly improves the accuracy and robustness of these critical regions, especially when the eyeball itself might not be perfectly convex.
  • Loss Function Exploration Matters: The choice of loss functions has a notable impact on segmentation quality, with boundary loss showing improvements in Mean Intersection over Union (MIOU) for the geometry head.
  • Eyelash Segmentation is Challenging but Important: While the noise head for eyelashes is crucial, the paper finds that modifying its loss function has less impact on overall performance compared to the geometry head, suggesting that the primary challenge lies in the annotation quality and the inherent variability of eyelashes.

Key Results

The proposed two-headed architecture achieves state-of-the-art results, with accuracy reaching 99% and precision/recall values over 0.98 for the geometry segmentation. Incorporating boundary loss improved MIOU from 0.9697 to 0.9751. The convex prior, when applied, significantly increased the iris convexity rate. Ablation studies showed that while boundary and surface losses benefit the geometry head, modifying the noise head's loss function had minimal impact on its quality.

Limitations and Future Work

The authors acknowledge that while the convex prior improves iris convexity, it might not be ideal for the entire eyeball, which isn't always convex. They suggest that enforcing convexity only on the iris and pupil is a more effective strategy. Future work includes exploring post-processing steps to further enhance representation quality for identification and investigating end-to-end learning of representations.

Summary updated

Abstract

Iris-based identification systems are among the most popular approaches for person identification. Such systems require good-quality segmentation modules that ideally identify the regions for different eye components. This paper introduces the new two-headed architecture, where the eye components and eyelashes are segmented using two separate decoding modules. Moreover, we investigate various training scenarios by adopting different training losses. Thanks to the two-headed approach, we were also able to examine the quality of the model with the convex prior, which enforces the convexity of the segmented shapes. We conducted an extensive evaluation of various learning scenarios on real-life conditions high-resolution near-infrared iris images.
BibTeX
@misc{lazarski2022two,
  author = {Wiktor Lazarski and Maciej Zieba and Tanguy Jeanneau and Tobias Zillig and Christian Brendel},
  title = {{Two-headed eye-segmentation approach for biometric identification}},
  year = {2022},
  eprint = {2209.15471},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2209.15471},
}
Figure