Transformer-based iris verification with attention-guided segmentation and Siamese learning

Area of Science:

• Biometrics and Pattern Recognition
• Computer Vision
• Artificial Intelligence

Background:

• Iris verification is a critical biometric technology for secure identification.
• Existing methods face challenges with image quality and complex iris textures.
• Accurate and efficient iris recognition systems are in high demand.

Purpose of the Study:

• To develop a unified framework for iris verification using deep learning.
• To enhance iris image quality and segmentation accuracy.
• To improve the discriminative power of iris feature representations.

Main Methods:

• A transformer-based Siamese network integrating image enhancement (denoising autoencoder) and attention-guided segmentation (U-Net).
• Global feature learning using a Vision Transformer to capture long-range dependencies in iris textures.
• Contrastive learning for optimizing feature discriminability.

Main Results:

• Achieved an Equal Error Rate (EER) of 2.34% and an Area Under the ROC Curve (AUC) of 0.987 on the CASIA-IrisV3 dataset.
• Demonstrated a True Acceptance Rate of ~95% at a False Acceptance Rate of 10⁻³.
• Attained 97.82% verification accuracy with an average inference latency of 12.6 ms.

Conclusions:

• The proposed integrated framework effectively enhances iris verification performance.
• Attention-guided segmentation and transformer-based features significantly improve accuracy and robustness.
• The method shows consistent performance across varied near-infrared acquisition conditions.