Robust Iris Segmentation Algorithm in Non-Cooperative Environments Using Interleaved Residual U-Net

Yung-Hui Li, Wenny Ramadha Putri, Muhammad Saqlain Aslam, Ching-Chun ChangSensors 2021 · OpenReview

TL;DR

This paper introduces a novel "Interleaved Residual U-Net" (IRUNet) for robust iris segmentation, crucial for accurate iris recognition, especially in challenging, non-cooperative environments. By combining deep learning for initial mask generation with a "Scan-and-Cluster" strategy for precise boundary refinement, the method achieves state-of-the-art performance on a difficult dataset.

What Problem Does This Paper Address?

Iris recognition is a highly accurate biometric identification method, but its effectiveness hinges on precise iris segmentation – accurately identifying the iris's inner and outer boundaries. Traditional methods struggle significantly in real-world scenarios where images are captured under "non-cooperative" conditions. These conditions include common issues like reflections (e.g., from glasses), partial occlusions (eyelashes, eyelids), off-axis viewing angles, motion blur, and low image resolution. These factors degrade the quality of iris images, making it difficult for algorithms to reliably extract the iris region, which in turn severely impacts the accuracy of subsequent iris recognition steps.

Methods

The proposed solution is a two-stage approach:

  1. Iris Mask Estimation (IRUNet): The core of the system is a custom-designed neural network called the "Interleaved Residual U-Net" (IRUNet). This network is based on the U-Net architecture, known for its effectiveness in image segmentation. The IRUNet incorporates:
  2. Interleaved Residual Blocks: These blocks are designed to combat the vanishing gradient problem common in deep networks and improve robustness against overfitting. They involve multiple convolutional layers with batch normalization and ReLU activations, with connections that help integrate features of different scales.
  3. Atrous Spatial Pyramid Pooling (ASPP): Inspired by Deeplabv3, this module uses dilated convolutions at different rates to capture multi-scale contextual information without losing resolution, which is crucial for segmenting fine details.
  4. The IRUNet performs semantic segmentation to generate an initial iris mask, essentially a binary image highlighting the iris region.
  5. Boundary Refinement (Scan-and-Cluster Strategy): After obtaining the initial mask, a novel "Scan-and-Cluster" strategy is employed to precisely locate the outer and inner iris boundaries.
  6. Outer Boundary: This involves scanning candidate points on the left and right edges of the iris mask within a specific vertical range (0.25x to 0.75x height of the ROI). Pairs of points are formed, and their distances are calculated. K-means clustering is then applied to these distances to identify the most likely boundary points, separating them from points affected by specular reflections or occlusions. The final outer boundary is fitted using these selected saliency points.
  7. Inner Boundary: A similar scanning approach is used, but adapted for the inner boundary. It involves identifying key endpoints and using scan lines to find intersection points with the iris mask. These points are then used to fit the inner boundary, aiming to accurately capture the pupil's edge.

Datasets

The primary dataset used for training and evaluation is the CASIA-Iris-Thousands database. This database contains 20,000 images from 1000 subjects, captured under near-infrared illumination. The authors also mention using UBIRIS.V2 for ground-truth mask generation. The ground-truth masks for both datasets were manually labeled using Labelme. The training/validation/testing split was 80:10:10.

Key Insights and Ideas

  • Deep Learning for Initial Segmentation, Traditional Methods for Refinement: The paper effectively combines the power of deep learning (IRUNet) for robust initial iris mask generation with a more traditional, geometric approach ("Scan-and-Cluster" with K-means) for precise boundary localization. This hybrid strategy leverages the strengths of both paradigms.
  • Addressing Non-Cooperative Environments: The "Scan-and-Cluster" strategy is specifically designed to handle challenges like specular reflections and occlusions by intelligently selecting saliency points, rather than relying solely on raw edge detection.
  • Interleaved Residual Blocks for Robustness: The novel IRUNet architecture with interleaved residual blocks is a key innovation for improving the network's ability to learn robust features and avoid common deep learning pitfalls like vanishing gradients and overfitting, which are critical for handling noisy real-world data.
  • Saliency Point Selection is Crucial: The paper highlights that simply detecting edges isn't enough. The intelligent selection of "saliency points" that are truly representative of the iris boundary is paramount for accurate segmentation, especially when dealing with occlusions and reflections.

Key Results

  • High Accuracy: The proposed method achieved mean IoU (Intersection over Union) values of 98.9% for outer boundary estimation and 97.7% for inner boundary estimation on the CASIA-Iris-Thousand database.
  • Outperforms State-of-the-Art: The results demonstrate superior performance compared to several existing methods, including conventional techniques like Circular Hough Transform and deep learning approaches like Faster-RCNN and U-Net, particularly on challenging images.
  • Efficient Execution Time: The proposed method achieved an execution time per image of 0.007 seconds, matching the efficiency of some of the fastest deep learning models tested.

Limitations and Future Work

  • Failure Cases: The authors acknowledge that the method can still fail in cases of very strong occlusion or when the initial predicted mask from the IRUNet is imperfect. In such scenarios, it becomes difficult to find enough reliable saliency points for accurate boundary recovery.
  • Future Work: The authors plan to explore strategies to make the IRUNet model more lightweight while maintaining its high accuracy and efficiency. This suggests a focus on model compression or architectural optimizations for deployment on resource-constrained devices.

Summary updated

Abstract

Iris segmentation plays an important and significant role in the iris recognition system. The prerequisite for accurate iris recognition is the correctness of iris segmentation. However, the efficiency and robustness of traditional iris segmentation methods are severely challenged in a non-cooperative environment because of unfavorable factors, for instance, occlusion, blur, low resolution, off-axis, motion, and specular reflections. All of the above factors seriously reduce the accuracy of iris segmentation. In this paper, we present a novel iris segmentation algorithm that localizes the outer and inner boundaries of the iris image. We propose a neural network model called “Interleaved Residual U-Net” (IRUNet) for semantic segmentation and iris mask synthesis. The K-means clustering is applied to select saliency points set in order to recover the outer boundary of the iris, whereas the inner border is recovered by selecting another set of saliency points on the inner side of the mask. Experimental results demonstrate that the proposed iris segmentation algorithm can achieve the mean IOU value of 98.9% and 97.7% for inner and outer boundary estimation, respectively, which outperforms the existing approaches on the challenging CASIA-Iris-Thousand database.
BibTeX
@inproceedings{li2021robust,
  author = {Yung-Hui Li and Wenny Ramadha Putri and Muhammad Saqlain Aslam and Ching-Chun Chang},
  title = {{Robust Iris Segmentation Algorithm in Non-Cooperative Environments Using Interleaved Residual U-Net}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=6OSpQ0dOTF},
}
Figure