A Novel Finger-vein Recognition Approach based on Vision Transformer
Huimin Lu, Yupeng Li, Chengcheng Zhao, Weiye Liu, Yang Li, Ning Ma · International Conference on Frontiers of Electronics, Information and Computation Technologies · 2021
As the second generation of Biometrics, finger vein recognition has been widely studied due to its advantages such as security, convenience, interior characteristics, and living body recognition. Deep learning-based recognition model represented by Convolutional Neural Network (CNN) has been applied in the field of finger vein recognition. Although CNN offers translational invariance, it provides limited rotational invariance by pooling. Inspired by the strong representation ability of the transformer in Natural Language Processing (NLP), we present a novel finger vein recognition approach based on Vision Transformer (ViT). To fit the Transformer architecture, we split finger-vein images into patches, linearly embed each of the patches, and fed the resulting sequence of vectors into the Transformer Encoder. The model uses multi-head attention mechanisms to obtain long-range contexture information in images without considering distances. We compared the performance of the proposed model with some typical models such as LeNet-5 and CNN-based model on three public finger-vein datasets. The experiment shows that our results achieved an average accuracy of 93.16%, 94.3%, and 92.11% on ViT-100, ViT-150, ViT-200 models, respectively. The results indicate that the new recognition method based on ViT is better than the existing models of feature representation.