Lightweight power line defect detection driven by channel decoupling and dynamic gating

Abstract

To address the challenges of complex backgrounds, tiny defect targets, and variable target scales in unmanned aerial vehicle (UAV) power inspection scenarios, this paper proposes an improved lightweight defect detection model based on YOLOv11n. First, a Channel Decoupled Representation Module is proposed to precisely focus on and extract fine-grained defect features while suppressing background noise in complex environments. Second, an Adaptive Gated Dynamic Upsampling operator is designed to replace traditional interpolation upsampling in the neck network. This maximally preserves the high-frequency detailed information of tiny targets and significantly reduces the missed detection rate. Then, a Shared Large Kernel Head is reconstructed to enhance cross-scale feature perception capability while maintaining the lightweight nature of the model. Finally, a Linear Threshold Dynamic Focusing Loss Function (LT-WIoU) is constructed to strengthen the learning of valid hard samples, which further improves the localization accuracy of the model. Experimental results demonstrate that, compared with YOLOv11n, the improved model achieves a 6.9% increase in mean Average Precision, alongside 0.8% and 7.9% improvements in precision and recall, respectively. Concurrently, the number of parameters and computational complexity are reduced by 16.6% and 22%, respectively. The proposed algorithm achieves an ideal balance between detection accuracy and lightweight design, providing a high-precision and highly reliable solution for intelligent dynamic UAV inspections.