RM-net: Uma Rede Neural Convolucional, com Função de Custo no Domínio da Frequência, para Realce de Imagens de Impressões Digitais Latentes.

Abstract

Fingerprint impressions play a crucial role in the veriőcation and biometric authentication of individuals. There are different types of őngerprints, including those captured in controlled environments using ink or scanners, and latent őngerprints, which are unintentionally left behind when őngers touch various surfaces. Enhancement procedures are commonly applied to improve the quality of ridges and facilitate the extraction of minutiae in őngerprints. However, studies have revealed that enhancement methods designed for inked or scanned őngerprints are not equally effective when applied to latent őngerprints. Currently, deep learning-based methods have been widely adopted in image processing. However, when attempting to reconstruct highly noisy parts of the images, these methods tend to őll in erroneously and non-naturally, or even amplify the existing noise in the original image, resulting in a worsening of the problem in some parts of the enhanced images. In view of this, the main objective of this research is to develop a speciőc enhancement method for latent őngerprint images that is capable of preserving ridge structures while seeking to mitigate or resolve this issue. The developed method is based on a convolutional encoder-decoder architecture, aiming to perform enhancement of latent őngerprint images. This architecture is designed to directly take the latent őngerprint image as input, without the need for preprocessing, and generate an enhanced version of the image as output. During the training of the method, a cost function was applied, which compares images in the frequency domain, resulting in a reduction of the problem of non-natural reconstruction of ridge structures in parts of the original image that exhibit high noise. Two experiments were conducted to validate the developed method: one related to latent-to-sensor matching and another related to latent-to-latent matching. The MOLF and IIIT-D latent őngerprint databases were used for these experiments. The obtained results demonstrated an improvement in őngerprint identiőcation and matching tests in the test set, when compared to scenarios where enhancement was not applied, as well as in comparison to methods found in the literature.