Shape-driven lightweight CNN for finger-vein biometrics

Finger-vein recognition has been widely explored for biometrics security over the past two or three decade. Its size and characteristics against presentation attacks make it suitable for various commercial, governmental and forensics applications. As with other biometric traits, finger-vein recognition approaches mainly depend on well-engineering feature descriptors and recently, more deep learning (DL) models have been designed to achieve high-performance biometrics. However, the existing DL approaches usually employ the neural networks with increasing layers and parameters, which inevitably leads to an increase in memory consumption and algorithm complexity. Besides, all the vein recognition networks use square filters, which are influenced by network designed for object detection and classification. In this paper, two network design issues for finger-vein recognition are explored: shallow network using rectangular filters and lightweight semi-pretrained network. The exhaustive experiments on three benchmark databases HKPU, FV-USM and SDUMLA demonstrate that rectangular filters outperform square filters for DL-based finger-vein recognition, and the lightweight semi-pretrained network also outperforms non-pretrained network.