Object Detection of Insulator and its Self-Explosive Defect Identification Based on Lightweight YOLOv5

To address the issue of large model size in existing transmission line insulator detection models, which limits flexible deployment, this study proposes a detection method for transmission line insulators and their self-explosive defects based on an improved You Only Look Once v5n (YOLOv5n). First, GhostNet_050 is used to replace the original backbone network Darknet53 of YOLOv5n, providing an initial step towards model lightweighting. Next, the small-object detection layer in the neck structure is removed, and the standard convolutions are replaced with Ghost convolutions, completing the lightweight optimization of YOLOv5n. The modified model is trained and validated on publicly available synthetic foggy insulator dataset (SFID), which contains insulators and their self-explosion defects. Experimental results show that, compared to the baseline YOLOv5n, the proposed modifications reduce the model's parameter count and computational load to 0.62M and 1.11 GFLOPS, representing decreases of 64.97% and 73.32%, respectively. Additionally, mean average precision (mAP) with an intersection over union (IoU) threshold of 0.5 only decreased by 0.06%, reaching 97.87%. This indicates a significant reduction in computational resources in specific detection scenarios with minimal impact on accuracy, making the model well-suited for real-time deployment in resource-constrained environments.