Uma nova abordagem baseada em grafos para o reconhecimento de impressões digitais infanto

Abstract

Infant fingerprint recognition presents significant challenges due to intrinsic skin characteristics, such as smaller areas and significantly thinner structures, and aging effect, which is a rapid developmental growth that induces morphological deformation over time. These factors render traditional recognition methods ineffective. This work introduces GBSEF, a novel graph-based framework designed to overcome limitations in existing approaches. The proposed methodology operates in an extractor-agnostic for minutiae manner, constructing graphs from minutiae data where nodes and edges are annotated with deterministically engineered attributes. A GNN then processes these structural representations to generate fixed-length embeddings. This final representation enables low-latency 1:1 and 1:N matching. Furthermore, the reliance on engineered—rather than learned—graph features provides inherent explainability, enabling direct traceability of matching decisions to human-interpretable fingerprint characteristics. Experimental evaluation on the longitudinal CMBD demonstrates that GBSEF substantially outperforms traditional open-source baselines, reducing the Equal Error Rate from 10.22% to 5.74% in zero-shot verification, for S2 subset. Employing a few-shot protocol with models trained in S1, improves performance to 2.99% and in cross-age that confirms temporal robustness, with a 2.43% on samples acquired four months post-enrollment. Thus, the framework establishes an efficient, reproducible, and transparent alternative for reliable longitudinal infant biometrics.