GAN-Based Image Augmentation for Finger-Vein Biometric Recognition

Deep learning methods, and especially convolutional neural networks (CNNs), have made a considerable breakthrough in various fields of machine vision, basically by employing large-scale labeled databases. However, deep learning methods applied in finger-vein area are basically implemented on small-scale datasets, which are probably faced with challenges such as overfitting, susceptible to finger position, unstable performance on various datasets and son on. In this study, we present a lightweight and fully convolutional Generative Adversarial Network (GAN) architecture, which is named FCGAN, using preliminary batch normalization, and tightly-constrained loss function for implementing finger-vein image augmentation. In addition, we present a novel scheme FCGAN-CNN for finger-vein classification, which reveals that synthetic finger-vein images using FCGAN are capable of improving the property of CNN for finger-vein image classification. The experiment of sample augmentation shows that the training accuracy using FCGAN-augmented samples could go beyond 99%, which is higher than 96.34% obtained using only classic sample augmentation. Furthermore, the well-trained CNN is further evaluated on a totally different dataset, which indicates that the proposed scheme FCGAN-CNN is capable of improving the ability of CNN to extract deep features. We consider that the proposed method for sample augmentation could be extended to other biometric systems.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.