A Finger Vein Liveness Detection System Based on Multi-Scale Spatial-Temporal Map and Light-ViT Model

Abstract

Prosthetic attack is a problem that must be prevented in current finger vein recognition applications. To solve this problem, a finger vein liveness detection system was established in this study. The system begins by capturing short-term static finger vein videos using uniform near-infrared lighting. Subsequently, it employs Gabor filters without a direct-current (DC) component for vein area segmentation. The vein area is then divided into blocks to compute a multi-scale spatial–temporal map (MSTmap), which facilitates the extraction of coarse liveness features. Finally, these features are trained for refinement and used to predict liveness detection results with the proposed Light Vision Transformer (Light-ViT) model, which is equipped with an enhanced Light-ViT backbone, meticulously designed by interleaving multiple MN blocks and Light-ViT blocks, ensuring improved performance in the task. This architecture effectively balances the learning of local image features, controls network parameter complexity, and substantially improves the accuracy of liveness detection. The accuracy of the Light-ViT model was verified to be 99.63% on a self-made living/prosthetic finger vein video dataset. This proposed system can also be directly applied to the finger vein recognition terminal after the model is made lightweight.