Detecção de equipamentos de subestação em imagem térmica através de Aprendizado Profundo

Abstract

Automatic detection of equipment in thermal images has proven to be a promising solution for improving inspections in electrical substations, making the process more efficient and reducing the need for human intervention. In this work, a comparative analysis was carried out among four modern object detection architectures — YOLO (You Only Look Once) in the YOLOv5 and YOLOv8 versions, Faster R-CNN (Region-Based Convolutional Neural Network), and RetinaNet — applied to a set of thermal images from electrical substations. The evaluation considered accuracy, generalization capability, visual performance, and training time. The models were trained and evaluated using performance metrics widely adopted in the literature, such as mean Average Precision (mAP), in addition to visual analysis of the detections and processing time. The results indicated that YOLOv8 achieved the best overall performance, reaching the highest mAP@50 and mAP@50–95 values, followed closely by YOLOv5, highlighting the robustness of the YOLO family for low-contrast thermal images. Faster R-CNN showed intermediate performance, while RetinaNet presented lower metrics and greater visual instability. Thus, the findings indicate that YOLO-based models are the most suitable for thermographic inspection applications, with potential for direct use in automated monitoring systems.