Aplicação de redes neurais convolucionais na extração de minúcias de impressões digitais de recém-nascidos

Abstract

Minutiae extraction in biometric fingerprint recognition, especially for children, represents a significant technical challenge. Although there are well-established extractors for adult fingerprints, many of them face challenges when applied to children’s fingerprints due to their smaller size and distinct characteristics. Among the available extractors, researchers of infant fingerprints recommend latent and high-resolution types, as they are typically effective in capturing details in lower quality and smaller fingerprints. Thus, this study aims to evaluate the effectiveness of MinutiaeNet, specialized in latent fingerprints, on images of children’s fingerprints. The adopted methodology includes the implementation of MinutiaeNet, data preprocessing, minutiae extraction via MindTCT, and the subsequent use of the Bozorth algorithm to calculate TAR, FAR, and CMC curve rates. The focus of the study is on performance comparison between the extractors to verify the consistency of MinutiaeNet’s accuracy. The results showed that MindTCT achieved a higher TAR compared to MinutiaeNet, with better performance in almost all FAR rates, especially at higher values. Additionally, the combination of minutiae extracted by both methods was explored. The analysis of the results showed that integrating the minutiae provided an increase in TAR, FAR, and the CMC curve.