Super-resolução para biometria de impressões digitais neonatais: um estudo sobre aprimoramento de imagens visando a equivalência com alta resolução

Abstract

Neonatal fingerprint recognition presents unique challenges due to the small size and low ridge definition of infant fingerprints, often requiring specialized acquisition equipment. As an alternative, super-resolution techniques have emerged as a promising solution to enhance fingerprint images captured at lower resolutions. In this study, we investigate whether deep learning–based super-resolution methods can faithfully reconstruct high-resolution fingerprint images from low-resolution inputs, and whether the resulting images exhibit comparable biometric performance to those acquired directly through high-resolution scanners. We evaluate three upscaling strategies—including a pre-trained model, a fine-tuned version tailored to neonatal data, and traditional bicubic interpolation—across multiple scaling factors. Our analysis considers visual similarity, biometric quality scores, and fingerprint matching performance using real high-resolution references as ground truth. The results indicate that super-resolution methods can produce visual and biometric characteristics comparable to those of high-resolution images, particularly at lower scaling factors. Additionally, we discuss potential limitations and raise questions about the reliability of structural fidelity in upscaled images, especially in sensitive biometric applications.