24 Feb
09:30 Master's Defense IC3 Auditorium
Minutiae Extraction in Non-Contact Fingerprint Images Using Deep Convolutional Networks
Anderson Nogueira Cotrim
Advisor / Teacher
Helio Pedrini
Brief summary
The area of ​​contactless fingerprint identification has proven to be a trend in recent years. An advantage of this acquisition category when compared to the capture of contact images is the high acceptance by users, as it is a less invasive technique and, in this pandemic period, avoids the use of a surface touched by other people. However, this area has several associated challenges. Contact sensors generally still produce greater biometric efficiency, as the minutiae are more highlighted due to the high contrast between the ridges and valleys. On the other hand, non-contact images usually have little contrast, causing methods to fail with spurious or undetectable minutiae, a characteristic that demonstrates the need for further studies in the area. In this work, we investigate the use of deep machine learning algorithms, given their success and the history of applicability of this type of approach in other research topics. Considering this information, we propose a method for minutiae extraction using convolutional neural networks. We developed and compared two architectures for minutiae extraction: WSMS-CNet, a patch-based approach that makes use of two-stage sharing for better multiscale robustness, and ResUSENet, an architecture for single-process prediction based on the U-shape and benefits from skip-connections and squeeze-and-excitation layers. In addition to the comparative study between both architectures, we propose a data augmentation method in order to avoid the presence of artifacts and reduce the overfitting effect. We also developed an algorithm for extracting minutiae from a probability map produced by the networks. The databases used in this work were acquired from the literature. For training, the FVC 2002 DB1A/DB3A, FVC 2004 DB1A/DB3A and CFPose data sets were used. For validation, CFPose, Benchmark 2D/3D and PolyU Cross datasets were used. Finally, the results showed that our method is competitive compared to state-of-the-art approaches and the commercial Verfinger tool.
Examination Board
Hélio Pedrini IC / UNICAMP
Pascual Jovino Figueroa Rivero Griaule SA
Marco Antonio Garcia de Carvalho FT / UNICAMP
André Santanchè IC / UNICAMP
David Menotti Gomes DInf / UFPR