Supervised Contrastive Learning and Intra-Dataset Adversarial Adaptation for Iris Segmentation.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Accurate iris segmentation is crucial for reliable iris recognition systems.
• Traditional methods struggle with non-ideal conditions (illumination, blur, squinting) and require extensive preprocessing.
• Deep learning methods are limited by small labeled datasets and fail to address distribution gaps in real-world iris data.

Purpose of the Study:

• To develop a robust iris segmentation method that overcomes limitations of small datasets and non-ideal conditions.
• To improve the performance of iris recognition by enhancing segmentation accuracy.
• To address the distribution gap within non-ideal iris datasets.

Main Methods:

• A three-stage training strategy combining supervised contrastive pretraining, cross-entropy fine-tuning, and intra-dataset adversarial adaptation.
• Supervised contrastive pretraining to enhance pixel classifier performance with limited data.
• Intra-dataset adversarial adaptation to align distributions of easy and hard samples in non-ideal irises.

Main Results:

• Significant improvements in iris segmentation performance across multiple datasets (UBIRIS.V2, IITD, MICHE-I, CASIA-D, CASIA-T).
• Achieved high F1 scores, including 96.66% on UBIRIS.V2 and 98.72% on IITD.
• Demonstrated effectiveness in handling non-ideal iris images with variations in illumination and motion blur.

Conclusions:

• The proposed three-stage training strategy effectively enhances iris segmentation accuracy, particularly for non-ideal conditions.
• The method successfully mitigates the impact of limited labeled data and distribution gaps.
• This approach offers a promising solution for more robust and accurate iris recognition systems.