Double-Center-Based Iris Localization and Segmentation in Cooperative Environment with Visible Illumination

Jiangang Li, Xin FengSensors 2023 · DOI

TL;DR

This paper introduces ICSNet, a novel end-to-end deep learning model for iris localization and segmentation that directly predicts iris boundaries from feature maps, bypassing traditional post-processing steps. This approach significantly improves robustness and accuracy, especially for iris images captured in challenging, non-cooperative environments with visible illumination.

What Problem Does This Paper Address?

Iris recognition is a highly accurate biometric technology, but its effectiveness hinges on robust preprocessing, specifically iris localization and segmentation. These steps are crucial for isolating the valid iris region from an eye image. However, iris images captured in "non-cooperative" environments (e.g., varying distances, visible lighting, user movement, and occlusions like glasses or eyelids) present significant challenges. Traditional methods, often relying on fitting circles to segmented iris masks using techniques like Daugman's integro-differential operator or circular Hough transforms, struggle with noise, blur, and irregular contours in these challenging images. This leads to inaccurate localization and segmentation, ultimately impacting the overall recognition performance. The paper aims to address the need for a more accurate, robust, and efficient method that can handle these real-world conditions without cumbersome post-processing.

Methods

The core innovation of this paper is the ICSNet (Iris Localization and Segmentation Network), an end-to-end deep learning framework that directly predicts iris boundaries from feature maps, rather than relying on intermediate masks.

Key methodological components include:

  • Double-Center Localization: Instead of fitting circles to masks, ICSNet treats the inner and outer iris circles as two distinct objects to be located. It uses a CNN backbone to generate feature maps, and then employs two sub-networks:
  • Center Localization Head: Predicts heatmaps indicating the probability of the inner and outer circle centers. This is inspired by heatmap-based keypoint detection.
  • Radius Regression Head: Regresses the radius of the inner and outer circles based on the predicted center points. An offset prediction is also included to correct for discretization errors.
  • Segmentation Module: A separate branch, integrated after the localization module, takes a Region of Interest (RoI) cropped using the predicted outer circle parameters and generates a precise iris mask.
  • Concentric Sampling Strategy (Auxiliary Training): To improve training on non-standard iris images with irrelevant facial features, a novel sampling strategy is introduced. This strategy leverages the natural biological characteristic of overlapping inner and outer iris circle centers to create a more focused training signal, particularly for the overlapping Gaussian distributions in the heatmaps. This strategy is cost-free during inference.
  • End-to-End Framework: The entire process, from input image to predicted circle parameters and iris mask, is performed within a single, unified network, eliminating the need for manual post-processing steps.

Datasets

The paper evaluates its proposed method on four publicly available iris datasets known for their challenging conditions:

  • NICE-II: Comprises two subsets (training and testing) with images captured in visible light.
  • MICHE: A collection of three sub-databases (GS4, IP5, and GT2), each with indoor and outdoor images captured using mobile devices. These datasets are particularly noted for their "non-cooperative" and noisy nature.

The paper also mentions using these datasets for training and testing, with specific image resolutions being adjusted to be multiples of 32 for the network's input.

Key Insights and Ideas

  • Feature Map Prediction is Key: The most significant insight is that directly predicting iris circle parameters from feature maps, rather than from intermediate masks, leads to greater robustness and accuracy, especially in challenging, noisy environments. This bypasses the limitations of mask-based fitting methods.
  • End-to-End Efficiency: Integrating localization and segmentation into a single network streamlines the process, reducing computational overhead and potential error propagation from separate modules.
  • Auxiliary Training for Real-World Data: The concentric sampling strategy highlights the importance of tailored training techniques for datasets that mimic real-world scenarios, where irrelevant facial features can distract standard models.
  • Biological Inspiration for Robustness: Leveraging the concentric nature of iris circles for sampling during training demonstrates how understanding biological characteristics can lead to more effective algorithmic solutions.

Key Results

The proposed ICSNet demonstrates strong performance across the evaluated datasets:

  • NICE-II: Achieved 84.02% box IoU and 89.15% mask IoU, outperforming existing methods.
  • MICHE (all sub-datasets): Achieved an average 74.06% box IoU, surpassing existing methods by 4.64%.
  • Comparison with State-of-the-Art: ICSNet consistently outperformed other methods (FCEDN, IrisDenseNet, FRED-Net, IrisParseNet, NIR-Zhang) in both localization and segmentation metrics across multiple challenging datasets, particularly the noisy MICHE databases.
  • Efficiency: The model is noted for being well-balanced in terms of accuracy and speed, requiring only 47.01 GFLOPs for processing (at 320x320 resolution).
  • Qualitative Results: Visualizations show ICSNet successfully localizing and segmenting irises in images where other methods failed, particularly those with significant noise, reflections, and occlusions.

Limitations and Future Work

The authors acknowledge several limitations and propose future research directions:

  • Serial Processing Time: The current network design processes localization and segmentation serially, which increases overall processing time.
  • Radius Regression Loss: The L1 loss for radius regression doesn't fully account for the paired nature of inner and outer iris radii.
  • Segmentation Improvement: While strong, the segmentation performance on some datasets could still be improved.

Future work will focus on:

  • Parallel Processing: Exploring more efficient network structures to perform iris localization and segmentation in parallel, reducing total processing time.
  • Optimized Radius Regression: Investigating alternative loss functions for radius regression that better capture the relationship between inner and outer radii.
  • Enhanced Segmentation: Developing more advanced techniques for improving segmentation accuracy on challenging, irregular iris images.
  • More Efficient Approaches: Exploring more efficient methods for improving segmentation performance, especially for irregular iris images captured under visible illumination.

Summary updated

BibTeX
@article{li2023double,
  author = {Jiangang Li and Xin Feng},
  title = {{Double-Center-Based Iris Localization and Segmentation in Cooperative Environment with Visible Illumination}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23042238},
}
Figure