Hand vein-based biometric authentication using two-channel similarity measure networks

In this paper the feasibility of convolutional neural networks (CNNs) for the purpose of authenticating an individual based on near infra-red (NIR) images of his/her dorsal hand vein patterns is investigated. In particular, a subset of CNNs called two-channel similarity measure networks (2CH-SMNs) is considered, since the problem at hand involves authentication (verification), instead of recognition (classification). All hand vein images are cropped and binarised before presented to the 2CH-SMN, in order to ensure that the focus of the network is on the structure of the hand veins, and not on the shape of the hand or on the grey-scale background variations surrounding the hand vein structure. A tailor-made 2CH-SMN is trained for each client enrolled into the system by considering authentic hand vein images from the client in question, as well as fraudulent images from other enrolled clients. Different images from the client in question and images from other non-training clients are employed for validation purposes, as well as for selecting optimal client-specific probability thresholds. Another distinct subset of authentic images, as well as images from imposters (non-clients) are used for testing. The results are encouraging.