Weather-Robust Insulator Detection for Smart Power Grids: An Enhanced DETR with Dynamic Attention and Scale-Aware Learning in UAV Inspection

This paper presents an optimized DETR(Detection Transformer) model designed to enhance the detection of small targets, such as insulators, under adverse weather conditions. As the demand for electricity grows, the safety of power grid systems becomes increasingly critical. Insulators, key components within these systems, must maintain functionality in harsh environments. Traditional detection methods are inefficient, and while drone technology has improved the level of intelligence, capturing highquality images in adverse weather remains challenging. This article improves the DETR model by incorporating an enhanced multi-scale feature extraction pyramid, a hybrid local-global attention mechanism, and a customized loss function. Experiments demonstrate that the IDD-DETR model surpasses other benchmark models in detection precision and recall under complex weather conditions. Compared to the YOLO series models, IDD-DETR achieves $17.3 \%$ higher average precision and excels in identifying small defects. Tests across multiple datasets also confirm its robustness.