Mixed Automatic Adversarial Augmentation Network for Finger-Vein Recognition
Yantao Li, Congyi Tang, Shaojiang Deng, Huafeng Qin, Hongyu Huang · IEEE Transactions on Instrumentation and Measurement · 2024
In recent years, vein recognition has attracted increasing attention as a secure and private biometric identification method. Despite the progress made by deep neural networks (DNNs) in vein recognition, existing solutions still suffer from limited robustness due to the lack of training image samples. To address this challenge, we propose MAdAugment, a Mixed Automatic Adversarial Augmentation network for finger-vein recognition, comprising a mixed augmentation network and a classifier trained adversarially. MAdAugment generates diverse samples to train a more robust vein classifier for finger-vein recognition, by optimizing the classifier and the augmentation network alternatively. Specifically, we first explore a mixed augmentation network to produce diverse finger-vein images, which consists of a policy network and a neural network. The policy network finds an optimal augmentation policy within the traditional augmentation operation search space, while the neural network refine augmented images with slight variations by the policy network. Then, we propose a novel adversarial loss incorporating the cosine similarity to reduce the search space. Finally, we conduct extensive experiments to evaluate the proposed MAdAugment on our and two public finger-vein datasets, and the experimental results demonstrate that MAdAugment outperforms existing approaches and effectively improves the recognition accuracy of the vein classifiers.