IALF-YOLO: Insulator defect detection method combining improved attention mechanism and lightweight feature fusion network

Abstract

Effect and efficient insulator defect detection is critical for advancing smart grid technologies. Current deep learning-based methods face limitations in small-object recognition accuracy, insufficient key feature extraction, and high computational complexity, which restrict their application in grid inspections. We propose IALF-YOLO, an improved YOLOv5s-based model, to address these problems. Firstly, by fusing the shallow feature map of Backbone and the deep feature map of Neck, a detection layer dedicated to small objects is created in Head, significantly improving small objects’ detection accuracy. Secondly, a S-CBAM attention mechanism is proposed, which addresses the issue of feature information loss in conventional CBAM by synchronizing the extraction channel with spatial attention. Finally, the lightweight GSConv module replaces the convolutional layer in the Neck network to construct a lightweight feature fusion network, which improves detection accuracy while reducing model complexity and the number of parameters. Our method improves mAP by 2.8% and 2.5% on both datasets, respectively. The detection speed is 2$\times$ faster than other methods.