Exploring Simple Siamese Networks for Finger Vein Authentication

Finger vein recognition has emerged as a reliable biometric authentication technology combining low forgery risk, high accuracy and low cost. However, challenges inherent to computer vision applications, such as inconsistent image quality and lighting condition variations persist, leading to the adoption of deep learning approaches for improved results. This study explores the use of Simple Siamese Networks to address the issue of reliance on inter-class comparisons, which can introduce bias and increase training complexity due to the need for careful sampling strategies. These issues are inherent in both traditional Siamese networks (which require negative or hard negative sampling) and classification based methods (which rely on the definition of decision boundaries). Originally developed for unsupervised learning, Simple Siamese Networks exclusively utilize positive image pairs to generate embedding representations that cluster images of the same class. In this study, augmented images from multiple shots of the same individual's finger are used as positive pairs, allowing the model to capture intra-class variability and generalize well without relying on inter-class comparisons. Extensive experiments were conducted on two publicly available datasets (FVUSM, MMCBNU), following a cross-validation protocol using three well-known Convolutional Neural Network (CNN) encoders: ResNet18, DenseNet121 and EfficientNet-B0. The proposed method achieves notably low Equal Error Rates (EER), with results as low as 0.056% on the FVUSM dataset. Our findings show that Simple Siamese Networks offer a robust, easier-to-tune and highly reliable alternative to existing approaches for finger vein recognition, eliminating the need for hard negative mining and class-based contrastive learning.