Fingerprint image enhancement using fully convolutional deep autoencoders / Destaque de imagens de impressão digital utilizando autoencoders profundos totalmente convolucionais

Image quality for fingerprint samples is critical for the matching process. Novel methods introduce deep learning matching techniques based on convolutions neural networks to enhance degraded fingerprint images. However, due to the nature of the enhanced image problem, these methods tend to rely on...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Brazilian Journal of Development 2022-05, Vol.8 (5), p.40027-40042
Hauptverfasser: Neto, Sandoval Veríssimo de Sousa, Batista, Leonardo Vidal, Guimarães, Pedro Ivo Aragão, Souza, Túlio Emanuel Santana de
Format: Artikel
Sprache:eng
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Image quality for fingerprint samples is critical for the matching process. Novel methods introduce deep learning matching techniques based on convolutions neural networks to enhance degraded fingerprint images. However, due to the nature of the enhanced image problem, these methods tend to rely on processing small image patches to achieve their goal. Such an approach may often yield satisfactory results while having high computational costs due to overlapping in patches. In this paper, we propose a fast and accurate fully convolutional neural network based on an auto-encoder architecture to enhance the quality of fingerprint images. We do not use the patch processing method and instead train a model to enhance the image as a whole. After exhaustive testing, we achieve a model that can quickly perform the desired task, while achieving an average of 97.956% and 83.748% per pixel accuracy on the easiest and hardest dataset respectively. The models were trained on the publicly available Fingerprint Verification Competition datasets. We then highlight the most general model that can best enhance the quality of all datasets.
ISSN:2525-8761
2525-8761
DOI:10.34117/bjdv8n5-474