Konferenzbeitrag
Exploring the Untapped Potential of Unsupervised Representation Learning for Training Set Agnostic Finger Vein Recognition
Vorschaubild nicht verfügbar
Volltext URI
Dokumententyp
Text/Conference Paper
Dateien
Zusatzinformation
Datum
2023
Autor:innen
Zeitschriftentitel
ISSN der Zeitschrift
Bandtitel
Quelle
Verlag
Gesellschaft für Informatik e.V.
Zusammenfassung
Finger vein patterns are a promising biometric trait because of their higher privacy and security features compared to face and finger prints. Finger vein recognition methods have been researched extensively, especially deep learning based methods such as Convolutional Neural Networks. These methods show promising recognition performance, but their low degree of generalization and adaptability results in much lower and inconsistent recognition performance in cross database scenarios. Despite these drawbacks, much less research has gone into the generalization and adaptability of these deep learning methods.
This study addresses these issues and proposes an unsupervised learning approach, namely a patch-based Convolutional Auto-encoder for learning finger vein representations. Our proposed approach outperforms traditional baseline finger recognition methods on the UTFVP, SDUMLA-HMT, and PKU datasets, and achieves state-of-the-art performance on the UTFVP dataset with 0.24\% EER. It also indicates a noticeably higher generalization of finger vein features across different datasets compared to a supervised method. The findings of this work offer promising advancements in achieving robust finger vein recognition in real-life scenarios, due to the enhanced generalization and adaptability of our proposed method.