Feature-Fused Deep Convolutional Neural Network for Dorsal Hand Vein Recognition

Dorsal hand vein recognition has attracted more and more attention from researchers due to its advantages of high recognition accuracy and good anti-attack performance. However, in practical applications, it is inevitably affected by certain external environments and bring out performance reduction, such as the droplet problem, which is rarely solved in current research works nevertheless. Facing this challenge, this paper proposes a feature-fused dorsal hand vein recognition model. Firstly, both dorsal hand vein matching and classification tasks are constructed via typical methods. Then, we introduce another classification task to learn the droplet and non-droplet features. Finally, the output feature vector of the droplet classification task is merged into other two tasks, meanwhile all the tasks are jointly optimized for the core purpose of promoting the performance of the dorsal hand vein matching task. The experimental result on our self-built dataset shows that the poposed model reaches 99.43% recognition accuracy and 0.563% EER, which achieves significant performance improvement in EER metric compared with the typical model.