RDA-YOLO: A robust dynamic adaptive network for tiny insulator defect detection

Insulator defect detection is a critical component in ensuring the safe operation of smart grids. To achieve more effective detection, image-based inspection utilising drone aerial photography offers advantages such as low cost, high efficiency, and superior accuracy. Compared to other approaches, the You Only Look Once (YOLO) method demonstrates outstanding performance in insulator defect detection. However, it struggles to achieve satisfactory results when detecting small defects against complex backgrounds. To address this issue, this paper proposes a high-precision insulator defect detection algorithm named RDA-YOLO, which builds upon the YOLOv8 algorithm as its baseline model. Firstly, a reverse large-selection kernel module is designed to effectively adjust the receptive field size, enhancing feature extraction capabilities for long insulator strings and minute features. Secondly, a Dynamic Head replaces the original detection head, utilising its unified attention mechanism to obtain more consistent classification and localisation features. Finally, a distribution-aware Wise-IoU metric is proposed, modelling bounding boxes as two-dimensional Gaussian distributions. By employing normalised Wasserstein distance, this enhances the network's detection capability for small targets. Experiments on a proprietary dataset demonstrate that, with only a modest increase in computational overhead, this network achieves 91.6% precision and 91.4% mAP0.5, outperforming other state-of-the-art algorithms. Moreover, we conducted extensive robustness experiments, which demonstrated that our approach achieves significantly enhanced robustness compared to baseline models, rendering it more suitable for detecting extreme weather conditions.