FVGNN: A Novel GNN to Finger Vein Recognition from Limited Training Data

Abstract

One cannot make bricks without straw, although deep learning has been widely used, it is a data hungry technique that requires numerous labeled samples. Unfortunately, finger vein dataset has a few images per class which is far from meeting the requirements. To alleviate this problem, considering the powerful ability of graph-based models on relational tasks, we innovatively propose an end-to-end graph neural network(GNN) FVGNN. Images are mapped into embedding node features and then concatenated with labels as inputs. The model learns how to compare inputs, rather than memorize a specific mapping from images to classes. Most of the previous algorithms has a lot of preprocessing and parameter tuning, but in our framework, these are not required. We test our lightweight framework on two well-known datasets, it converges quickly and gets promising results with the accuracy of 99.98%, which outperforms the previous best result.