A finger vein authentication method based on the lightweight Siamese network with the self-attention mechanism

Finger vein authentication is an efficient and convenient biometric verification technology which has been widely investigated in personal identification. However, the vein image quality can be reduced by changeable finger postures and near-infrared light distribution during the finger vein collection procedure. The poor finger vein images tend to cause feature loss, affecting final authentication results. Therefore, we propose a lightweight Siamese network with a self-attention mechanism to improve the authentication performance for low-quality finger vein images. Firstly, the finger vein features are extracted from a three-layer convolutional neural network. Then, we introduce the global context network to model the feature maps from different layers, obtaining global information about the vein features. Next, a multi-scale feature fusion strategy is proposed to fuse the feature maps from different scales, which increases the diversity of vein features. Finally, the self-attention convolution module is proposed to weight and vectorize the fused features. Extensive experiments are conducted on three datasets, MMCBNU_6000, SDUMLA-HMT, and FV_USM, demonstrating that the proposed method can significantly improve over the state-of-the-art methods.