Aplicando super-resolucão para melhorar a extração de características em impressões digitais de recém-nascidos

Abstract

One of the limitations of newborn fingerprint recognition systems, compared to those for adults, is the smaller size of the fingers, which results in smaller image captures. This causes conventional biometric comparators, which generally rely on the distance between fingerprint features, to not function correctly for newborn biometrics detection. To solve this problem, one option is to use high-resolution sensors, which are costly. Another alternative is to perform computational resizing of the image, generally using bilinear or bicubic interpolation. However, these resizing methods can introduce noise and degradations such as image blurring, which hinder the recognition of minutiae, and exacerbate other degradations caused by the condition of the fingers. In this context, this work focuses on applying Super-Resolution networks, with the aim of resizing the image and removing possible noise and degradations. Super-Resolution is applied as a form of pre-processing the fingerprint, which is subsequently segmented, and its minutiae are extracted and compared with another capture. To evaluate improvements in biometric verification (1:1), TAR and FAR are measured for different thresholds, shown in the form of ROC curves. Similarly, for biometric identification, it is verified whether the fingerprint belongs to the set of the t-most similar prints, displayed in the form of CMC curves. Based on the results obtained, visual improvements in the images were observed, as well as an increase of up to 9.7% in accuracy in biometric verification and 8.58% in biometric identification at certain resolutions.