A Rapid Finger Vein Recognition Approach Based on UNet++

With the advancement of biometric technologies, finger vein recognition has garnered widespread attention within the academic community. This paper proposes a method based on the UNet++ model, wherein parameter adjustments are made to the UNet++ architecture to ensure precise pixelwise segmentation of finger vein images. Following the parameter optimization, model compression is applied through structural reparameterization to further increase the recognition speed of the trained model. The approach achieves recognition accuracies of 93.1% on two publicly available finger vein datasets, and it features a feature extraction time of 0.0131 s for a single finger vein image. The recognition time for a single fingerprint vein image is 0.0294 s. The results indicate that the use of the UNet++ model as a framework for finger vein feature extraction and subsequent model compression render it suitable for finger vein recognition tasks.