Joint Attention Network for Finger Vein Authentication

Abstract

Finger vein, as an intrinsic trait, has received increasing attention from academia and industry for its liveness detection ability and high security. However, the principle of vein imaging and its contact-less mode results in low contrast of finger vein image, where especially the vein texture is susceptible to factors, such as scattering in skin and tissue, ambient illumination, and finger posture variation. To extract discriminative features from low contrast images, in this article, we propose an attention mechanism, namely, joint attention (JA) module, which enables dynamic adjustment and information aggregation in the spatial and channel dimensions of feature maps to focus on fine-grained details, thus enhancing the contribution of vein patterns to extract identity features. Besides, we embed a generalized mean (GeM) pooling layer into our network to reduce the dimensionality of feature maps and output a compact and highly expressive feature representation. Finally, we build a new finger vein authentication architecture, called JA finger vein network (JAFVNet), with competitive experimental results on multiple public and self-built datasets: 0.35% EER for SDUMLA-HMT, 0.08% EER for MMCBNU-6000, 0.34% EER for FV-USM, and 0.49% EER for FV-SCUT. Numerous ablation experiments further illustrate the effectiveness of each component in the network.