26 April 2024
14:00 Master's Defense Room 85 of IC2
Features Extended by Contrastive Learning in Person Identification from Latent Fingerprints
André Igor Nóbrega da Silva
Advisor / Teacher
Alexandre Xavier Falcao
Brief summary
Many individual recognition systems use body biometrics, with systems based on fingerprints standing out. This biometric trait is considered universal, distinct between different individuals, invariant throughout life and easy to measure. Therefore, it is widely used in applications such as forensic investigations, financial, health and social services, among others. There are 3 types of fingerprints: rolled, deposited and latent. The first two are obtained by sensors in controlled environments and have high image quality. Latents, on the other hand, are unintentionally left and captured by forensic agencies at crime scenes. The main algorithms for comparing different impressions use local structures known as minutiae as a form of representation. Minutiae are the centers of characteristic regions in the image, such as bifurcations or abrupt ends of lines. Algorithms based on details are very successful in applications with rolled or landed fingerprints, proven by the high rates of correct identifications in literature databases. However, they fail to deal with latents, given the poor overall image quality and partial finger capture. To overcome this problem, techniques for representing and comparing fingerprints based on Deep Learning have been proposed and present promising results in the area. Within this context, this dissertation's main objective is to improve the computational capacity of representing latent fingerprints through the use of extended minutiae characteristics, learned with contrastive learning; and the improvement of matching algorithms that exploit such characteristics. To this end, we propose the investigation of Contrastive Learning, data augmentation techniques and a weak annotation methodology to train Siamese Networks even with a very limited database. Furthermore, we intend to explore new matching algorithms that use both the geometric components of the minutiae and the deep description component, so that the comparison between fingerprints is made in a more robust way. We validate the proposed methodology through identification experiments on the challenging NIST SD27 database, widely explored in the literature. We show that there is a significant improvement (~10%) in the correct identification rate of individuals when combining extended features with traditional minutiae matching methods.
Examination Board
Alexandre Xavier Falcão IC / UNICAMP
David Menotti Gomes DInf / UFPR
Marcelo da Silva Reis IC / UNICAMP
Hélio Pedrini IC / UNICAMP
João Paulo Papa FC / UNESP