Vein-based Biometric Verification using Transfer Learning

Vein patterns have been nowadays established as a reliable and convenient solution to implement biometric recognition systems. Yet, there is still a relevant gap between the real potential of such modality and its current usage in state-of the- art commercial solutions. Actually, one of the major issue limiting the proper exploitation of deep learning approaches for vein-based biometric recognition is related to the moderate sample size of existing publicly available datasets, typically not large enough to perform a comprehensive training of dedicated networks. In order to mitigate the lack of training generalization, instead of proposing novel convolutional neural networks to be trained from scratch, we here investigate the usefulness of employing existing architectures, and applying transfer learning to design a highly-efficient biometric recognition system based on vein patterns. Verification performance in open-set scenario is evaluated applying the proposed approach to three different kinds of vein data, that is, finger veins, palm veins, and hand dorsal veins. Results exceeding current literature are obtained on the SDUMLA, PolyU, and Bosphorus vein datasets, with equal error rates in open-set verification conditions at 0.41%, 0.006%, and 5.63%, respectively.