A multibiometric system based on finger photo and palm photo

Abstract

Recognition of finger photos and palm photos is an emerging area in biometrics, primarily leveraging textural attributes for authentication. Limited efforts have explored the application of deep neural networks (DNNs) for finger photo minutiae extraction, while palm photo minutiae extraction remains challenging due to issues such as blurriness and creases, which can lead to false minutiae. In this paper, we introduce a generative adversarial network (GAN) that utilizes frequency-domain patches to deblur and enhance image quality, effectively addressing both deblurring and crease removal. Furthermore, we present a deep convolutional neural network (DCNN) for minutiae extraction from enhanced patches, alongside a model for singular point (SP) detection. Additionally, we propose a score-based multibiometric system that seamlessly integrates palm and finger photos without the need for score normalization. To validate the effectiveness of our approach, we conducted extensive experiments on a database comprising 30,000 hand photos from 2,500 volunteers, as well as a smaller dataset of 2,400 hand photos from 200 volunteers for cross-database evaluation. Our results demonstrate the enhanced accuracy of our models and establish the superiority of our multibiometric system over state-of-the-art (SOTA) methods.