An insulator defect detection approach based on improved feature pyramid and attention mechanism

Insulators are crucial components of transmission lines for maintaining the normal operation of the power grid. With the improvement of UAV inspection equipment, the resolution of photos has significantly improved. However, these photos often contain multiple insulator strings of different sizes, making it difficult for traditional target detection algorithms to efficiently and accurately detect all defects such as Insulator self-explosion. To address this problem, we propose an improved target detection algorithm suitable for insulator defect detection. Our algorithm is based on the YOLOv5 detection network, which uses a bidirectional feature pyramid network to improve feature fusion capability. We also integrate feature extraction with ECANet attention mechanism and use Soft-NMS to prevent missing detections of close insulator strings, thus improving the comprehensive detection capability of insulator strings. We compared our proposed algorithm with the original YOLOv5 network using collected patrol images. Our results show that the improved algorithm increases mAP_0.5 and mAP_0.5:0.95 precision mean value indicators by 2.9% and 1.7% respectively.