A Lightweight Detection Method for Insulators in Transmission Lines Based on Improved YOLOv5

To address the issue that traditional neural network image recognition models cannot be effectively deployed on micro-embedded devices such as drones, a lightweight insulator image recognition algorithm for power inspection on drone-embedded platforms is proposed based on the YOLOv5 algorithm. Firstly, this paper replaces the original backbone network CSPDarknet53 with the lightweight module MobileNetV3A to improve detection accuracy while reducing computational costs. Secondly, the presented CBSA attention mechanism aims to strengthen the model’s capacity to concentrate on targets. Finally, an improved pyramid pooling module (SimSPPFC) is added after the backbone network to enable more effective fusion of local and global features. The experimental outcomes demonstrate that the optimized lightweight YOLOv5 algorithm reaches a mAP of $98.01 \%$, reduces the model size by $78 \%$, and decreases computational cost by $89 \%$. It meets the deployment requirements of edge devices while ensuring high detection accuracy.