Multi-scale feature fusion network optimization and edge intelligent detection for transmission line insulator

In response to the accuracy loss caused by lightweight networks on edge devices and the difficulty in accurately detecting small defect targets, this paper studies the lightweight You Only Look Once version 8 (YOLOv8) transmission line insulator defect detection algorithm based on Three-Branch Attention Mechanism (TBAM) and Enhanced Multi-Scale Feature Fusion Network (EMSFFN), which can perform high-precision detection on edge intelligent devices of unmanned aerial vehicles. Aiming at the problem of significant differences in the number of image categories in the current aerial insulator defect dataset, the dataset was expanded through image data augmentation to generate more minority class samples, balance the number of samples in each category. Experiments show that compared with YOLOv7, YOLOv8, YOLOv9, Detection Transformer (DETR) and Faster Region-based Convolutional Neural Network (Faster R-CNN), the model parameters and computational complexity of TBAM-YOLOv8-EMSFFN are relatively low. Its AP reaches 99.2% and the mAP reaches 98.7%. Therefore, the TBAM-YOLOv8-EMSFFN model is suitable for unmanned aerial vehicles carrying edge intelligent devices to achieve power inspection.