Uma Abordagem em CNN para Filtragem Rápida de Minúcias em Fingerprints

Abstract

Minutiae extraction represents a fundamental process for ensuring reliability in fingerprint recognition systems. This paper introduces an optimi zed hybrid architecture combining adaptive Gabor filtering with an extremely lightweight CNN (125,601 parameters) for robust minutiae validation. Our CNN post-processing stage analyzes 29×29 pixel patches from the initial Gabor iteration, achieving state-of-the-art performance on NIST SD4 with a 1.89% Equal Error Rate- a 30% improvement over conventional approaches. The implementation demonstrates remarkable efficiency, processing each minutia in 1ms through optimized C++ single-thread execution. These advancements prove that carefully designed lightweight neural architectures can significantly enhance biometric systems while meeting real-time operational constraints in embedded environments.