Detection of suspicious facial objects in neutral ATMs using deep learning architectures based on YOLOv8 and faster R-CNN

Aragon Paucar M.M.; Fernandez Acero K.Y.; Sulla Espinoza E. · Applied Computer Science (2026) · DOI: 10.35784/acs_8906

Resumen

This study presents an automated detection system for suspicious facial objects in neutral automated teller machines (ATMs) using two deep learning-based detection architectures: YOLOv8 and Faster R-CNN. The dataset was constructed from videos captured in ATM environments and complementary images associated with facial occlusion. From this material, frames corresponding to three classes (masks, hats, and sunglasses) were extracted, filtered, and annotated. Both models were trained using transfer learning and evaluated using standard object detection metrics, including precision, recall, F1-score, IoU, mAP@0.5, and mAP@0.5:0.95. The results show distinct behaviors between the two architectures. YOLOv8 achieved higher precision across the evaluated classes, reducing false positives and providing a more stable response to avoid unnecessary alerts. In contrast, Faster R-CNN achieved higher recall, showing greater sensitivity to partially visible facial objects, though with a higher tendency to generate false detections. Additionally, a deterministic rule was incorporated to classify each scene as NORMAL or SUSPECT, based on the number of valid suspicious classes detected and their temporal persistence in video sequences. The proposed system does not perform biometric identity recognition but instead focuses on identifying visual conditions associated with facial concealment. The findings suggest that deep learning-based object detection can support security in neutral ATMs through early visual alerts, although further validation in real operational environments is still required.

Instituciones: Universidad Nacional de San Agustín de Arequipa