Multi-Scale Arc-Fusion Based Feature Embedding for Small-Scale Biometrics

Automatic person recognition using finger-vein has been extensively investigated in recent years. Mostly the published research on biometrics are based on the hand-crafted feature descriptors to describe the appearance of a finger-vein image, which is prone to be affected by irregular shading, irrelevant background, varying illumination and non-rigid deformation. To improve the recognition performance of hand-crafted feature-based methods, many researchers have contributed substantially in designing enhancement methods for texture descriptors. However, hand-crafted algorithms are highly targeted with weak generalization ability with respect to the emerging data samples. Nowadays, deep convolutional neural networks (CNN) is emerging as a powerful technology to extract multi-level feature representation instead of traditionally designed features from raw data. These methods are achieving unprecedented performance in the field of computer vision. In context to biometrics modalities, finger-vein recognition using CNN is still in its primary stage. In this paper, we proposed a simple yet effective multi-scale arc-fusion approach to optimize the feature embedding discrimination power. An exhaustive comparative experiment is conducted on four publicly available databases: HKPU, FV-USM, SDUMLA and UTFVP, to demonstrate that the proposed approach is superior to the existing state-of-the-art (SOTA) methods. Also, the experimental result shows that the concept is actually scalable and can be migrated to other deep neural network architectures.