Finger Vein Identification Using Pretrained Feature Matching Networks

This work investigates the feasibility of using pretrained, feature-based matching neural networks for finger-vein–based person identification without retraining on biometric datasets. The proposed solution compares two finger-vein images by extracting point correspondences and computing a similarity measure based on the number and spatial distribution of matches. To enforce geometric consistency, we apply homography-based verification using the matched points. We evaluate the capability of several pretrained neural network models (SuperGlue, GlueStick, ASpanFormer, LoFTR, and SGM-Net) to verify and identify individuals based on images of finger veins on three publicly available datasets (SDUMLA, MMCBNU, and FV-USM) without additional training or fine-tuning. Experiments cover verification via pairwise comparisons and open-set identification, using a single parameter setting across all datasets to assess robustness. Verification achieves an accuracy above 99%. In the open-set identification setting, the best result yields an equal error rate (EER) below 3.5%. These results indicate that general-purpose matching networks can transfer effectively to finger-vein recognition without biometric-specific retraining.