U-Net, LW-Net and Siamese Network in Retinal Biometric Recognition

This paper presents a comprehensive study on retinal biometrics, highlighting the effectiveness of retinal vessel patterns as reliable biometric identifiers due to their unique and stable characteristics. We provide an extensive review of retinal anatomy, publicly available datasets, and current state-of-the-art segmentation and verification methods. The study emphasizes the importance of accurate vessel segmentation, evaluating U-Net and LW-Net neural networks, and investigates identity verification performance using Siamese networks based on a ResNet50 backbone with contrastive loss. The final configuration achieved an Area Under the Curve (AUC) of 0.9792, an F1-score of 90.8%, and a False Positive Rate (FPR) of 7.1%.