Finger Vein Recognition Using a Shallow Convolutional Neural Network

Deep learning-based finger vein recognition (FVR) can be classified as either a closed-set architecture (CS-architecture) or an open-set architecture (OS-architecture) based on the system output. The CS-architecture has limited practicality due to its closure, and the OS-architecture has limited generalization ability due to its challenging convergence. To improve the practicality and performance of deep learning-based FVR, we hypothesize that a shallow convolutional neural network is suitable for FVR based on the observation of the difference between face recognition and FVR. Consequently, we design a shallow network with three convolutional blocks and two fully connected layers that can be efficiently applied for both CS-architecture and OS-architecture. Moreover, an improved interval-based loss function is used to extract discriminative large-margin features. The proposed network has excellent performance, verified by extensive experiments on publicly available databases.