Insulator Defect Detection Based on Lightweight Network and Enhanced Multi-scale Feature Fusion

With the development of target detection algorithm embedded in UAV for insulator inspection of transmission towers, a YOLOv5-3S-4PH model based on lightweight network and enhanced multi-scale feature fusion is proposed to detect insulator defects in real time in view of the low detection speed, high network complexity and the difficulty of accurate detection of small defect targets. Firstly, the reconstructed ShuffleNetV2-Stem-SPP(3S) network is used as the backbone of YOLOv5, which reduces the amount of network parameters and calculation significantly. Secondly, the enhanced multi-scale feature fusion network for small targets and four prediction heads(4PH) is added to enhance the network's perception of insulator defects. Combined with Mosaic-9 data enhancement and CIoU loss function, the loss of detection accuracy caused by lightweight is further compensated. Finally, the YOLOv5-3S-4PH model is applied to the self-made insulator dataset for verification. The experimental results show that mean average precision(mAP) is increased by 3%, the detection speed is increased by 81.8%, and parameters and calculation are decreased by 82.4% and 67% compared to original YOLOv5 model. Therefore, the proposed model is more suitable for real-time monitoring of insulator defects deployed on UAV platforms.