SIAMESE FUSION U-NET FOR FINGER VEIN BIOMETRIC RECOGNITION

Abstract

Finger vein recognition (FVR) is a promising biometric modality due to its resistance to spoofing, contactless usage, and lifelong stability. However, reliable recognition remains challenging because finger vein images often suffer from low contrast, light scattering, and noise introduced by biological tissues. This study proposes a novel deep learning-based framework that jointly enhances segmentation and classification performance through two key innovations. First, introduce Atten-D3Net, an enhanced U-Net variant that integrates dilated fused convolutional layers, depth-wise separable fused convolutions, and a spatial attention mechanism to capture fine-grained vein patterns under challenging imaging conditions. Second, develop the Siamese Cross-Folded Deep Fusion (S-CFDF) network, which leverages cross-attention and weighted loss functions to improve the discrimination of subtle inter-class vein variations while reducing misclassification. Preprocessing with CLAHE, histogram equalization, bilateral filtering, and sharpening further improves image clarity. Experimental evaluations on two public datasets (MMCBNU_6000 and FV) demonstrate the superiority of the proposed system, achieving accuracies of 97.5% and 98.5%, respectively, while reducing computation time by more than 50% compared to existing baselines. Performance is validated through precision, recall, F1-score, Dice coefficient, and Jaccard index, showing consistent improvements over VGG16, DenseNet, Vision Transformer, and EfficientNet. The results highlight that the combination of AttenD3Net and S-CFDF enables robust, accurate, and efficient finger vein recognition, offering strong potential for deployment in secure biometric authentication systems.