Robust Finger Vein Recognition Based on Lightweight Attention Convolutional Neural Networks

In the field of computer vision (CV), convolutional neural networks (CNN) have demonstrated superior capability in feature extraction. However, the translation during the process of capturing finger vein images can potentially degrade the accuracy rate of the model, making it challenging to apply CNN to real-time and high accuracy finger vein recognition. Additionally, while CNN offers high accuracy, they also require a high number of parameters. Previous studies also confirmed the absence of shift-invariant features in CNN. In the light of these issues, this research proposes a lightweight attention convolution neural network (LACNN) for applications in finger vein recognition. The model incorporates diverse branch block (DBB), adaptive polyphase sampling (APS), and coordinate attention mechanism (CAM). To evaluate the performance of the model under finger vein recognition, the finger-vein by university sains malaysia (FV-USM) public database is employed for analysis and comparison with recent research methods. According to the result, the new model achieves an accuracy recognition rate of 99.82% with a parameter count of only 1.23 million. Thus, the LACNN presented in this study surpasses the other recent finger vein recognition methods.