Percorrer por autor "Amaral, Ricardo Filipe Couceiro"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- Artificial Intelligence-Driven Cascaded Screening for Age-Based ImageTriagePublication . Amaral, Ricardo Filipe Couceiro; Negrão, Miguel Cerdeira Marreiros; Frade, Miguel Monteiro de Sousa; Domingues, Patrício RodriguesThe exponential growth of digital visual content associated with Child Sexual Abuse Material (CSAM) has transformed forensic investigations into large-scale triage prob lems, particularly in cases involving the identification of underage individuals. Con ventional age classification systems, designed under closed-set assumptions, are fun damentallyinadequateforthissetting: theyignoreuncertainty, produceoverconfident predictions, and fail to meet the safety requirements of forensic decision-making. This work introduces a risk-aware age screening framework that separates internal uncertainty modelling from final binary classification. The Primary Screening Model (PSM) applies a calibrated dual-threshold policy to identify high-confidence and am biguous cases. PSM-indeterminate cases are then routed to the Secondary Refinement Model (SRM), whichproduces a binary refinement decision. This design ensures that the complete system produces a final binary screening output for detected faces, elim inating model-level indeterminacy entirely. Outputs classified as POSSIBLE_MINOR are subject to mandatory humanconfirmationpriortoanylegallyconsequentialaction not because the pipeline is unresolved, but because the ethical and legal weight of de cisions involving potential child-age class demands human accountability at the point of consequence. The proposed framework integrates probabilistic calibration, risk-constrained threshold selection, explainability via Gradient-weighted Class Activation Mapping (Grad-CAM), and cryptographic traceability mechanisms, ensuring both technical robustness and forensic admissibility. By unifying calibrated uncertainty modelling, PSM-level routing, cascaded refine ment, and forensic operational constraints, this work moves beyond traditional age estimation and establishes a practical, scalable, and defensible approach to real-world forensic triage.