Authors:
Margarida Gouveia
1
;
2
;
Eduardo Castro
1
;
2
;
Ana Rebelo
1
;
Jaime Cardoso
1
;
2
and
Bruno Patrão
3
;
4
Affiliations:
1
Instituto de Engenharia de Sistemas e Computadores, Tecnologia e Ciência, Porto, Portugal
;
2
Faculdade de Engenharia, Universidade do Porto, Porto, Portugal
;
3
Imprensa Nacional-Casa da Moeda, Lisbon, Portugal
;
4
Department of Electrical and Computer Engineering, University of Coimbra, Coimbra, Portugal
Keyword(s):
Biometrics, Convolution Neural Network, Equivariance, Fingerprints, Group Convolutional Network, Minutiae, Multi-Task Learning, U-Net.
Abstract:
Currently, fingerprints are one of the most explored characteristics in biometric systems. These systems typically rely on minutiae extraction, a task highly dependent on image quality, orientation, and size of the fingerprint images. In this paper, a U-Net model capable of performing minutiae extraction is proposed (position, angle, and type). Based on this model, we explore two different ways of regularizing the model based on equivariance priors. First, we adapt the model architecture so that it becomes equivariant to rotations. Second, we use a multi-task learning approach in order to extract a more comprehensive set of information from the fingerprints (binary images, segmentation, frequencies, and orientation maps). The two approaches improved accuracy and generalization capability in comparison with the baseline model. On the 16 test datasets of the Fingerprint Verification Competition, we obtained an average Equal-Error Rate (EER) of 2.26, which was better than a well-optimiz
ed commercial product.
(More)