Abstract
Unsupervised learning methods can learn generalized features without label in-
formation, which can be advantageous for small datasets like finger veins. This
work explores the possibility of learning finger vein representations with an un-
supervised learning method called Convolutional Auto-encoder(CAE), and pro-
posed a modification to the loss function to learn better representations of the
vein patterns. The results indicate that the CAE is powerful in reconstructing
global structures, yet failed to reconstruct vein patterns. The CAE with a Log-
likelihood ratio classifier achieved 6.74% EER on UTFVP dataset. Though the
performance of the system is far from the state-of-the-art, the findings imply that
the global finger structures contribute to the identity, and are powerful enough
to be used as an additional information source for finger vein recognition.
formation, which can be advantageous for small datasets like finger veins. This
work explores the possibility of learning finger vein representations with an un-
supervised learning method called Convolutional Auto-encoder(CAE), and pro-
posed a modification to the loss function to learn better representations of the
vein patterns. The results indicate that the CAE is powerful in reconstructing
global structures, yet failed to reconstruct vein patterns. The CAE with a Log-
likelihood ratio classifier achieved 6.74% EER on UTFVP dataset. Though the
performance of the system is far from the state-of-the-art, the findings imply that
the global finger structures contribute to the identity, and are powerful enough
to be used as an additional information source for finger vein recognition.
Original language | English |
---|---|
Title of host publication | Proceedings of the 2021 Symposium on Information Theory and Signal Processing in the Benelux |
Place of Publication | Eindhoven, the Netherlands |
Publisher | Werkgemeenschap voor Informatie- en Communicatietheorie (WIC) |
Pages | 43-51 |
ISBN (Electronic) | 978-90-386-5318-1 |
Publication status | Published - 2021 |
Event | 41st Symposium on Information Theory and Signal Processing in the Benelux 2021 - Online Symposium Duration: 20 May 2021 → 21 May 2021 Conference number: 41 |
Conference
Conference | 41st Symposium on Information Theory and Signal Processing in the Benelux 2021 |
---|---|
City | Online Symposium |
Period | 20/05/21 → 21/05/21 |