Self-Supervised Learning Framework toward State-of-the-Art Iris Image Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Iris segmentation is crucial for iris recognition systems.
• Deep learning methods require extensive, high-quality labeled datasets for optimal performance.
• Generating sufficient labeled data for iris segmentation is a significant challenge.

Purpose of the Study:

• To develop a self-supervised framework for generating unlimited, diverse iris images.
• To train iris segmentation networks using synthetic data to achieve state-of-the-art performance.
• To propose a generalized framework applicable to various object segmentation tasks.

Main Methods:

• Utilized a pix2pix conditional adversarial network for generating synthetic iris images.
• Developed an algorithm with 11 tunable parameters for random iris mask generation.
• Employed generated images for training deep learning-based iris segmentation networks.

Main Results:

• Achieved state-of-the-art performance in iris segmentation using generated training data.
• Demonstrated promising results across all commonly used performance metrics.
• Validated the framework's ability to produce photo-realistic training data.

Conclusions:

• The proposed self-supervised framework effectively generates unlimited synthetic iris data.
• This method significantly improves iris segmentation network performance.
• The framework is adaptable for other object segmentation tasks with minor adjustments.