Answers are grounded in this paper and cite their sources.
Insulator defect detection plays a critical role in ensuring the reliable operation of power transmission systems. Traditional manual inspection remains labor-intensive and susceptible to subjective errors. Although deep learning-based object detection has demonstrated considerable potential in this domain, existing models continue to struggle with the accurate identification of small defects and the robust handling of complex backgrounds commonly encountered in aerial inspection imagery. To address background clutter and the coexistence of large insulator bodies and small defect regions, we adopt the Reparameterized Hierarchical Multi-Scale (RepHMS) module and extend its spatial branches with learnable weights. These weights adjust the relative contributions of local detail and wider context and can be folded into the deployment convolution. We also adapt Multi-Scale Gradient Descent (MuSGD) to the depthwise convolutions in RepHMS by applying Newton–Schulz orthogonalization to each channel’s spatial kernel and matching the layer-level update magnitude. Experiments on an insulator defect dataset show our method achieves 99.3% mAP@0.5 and 87.7% mAP@0.5:0.95, outperforming baseline by 4.4% and 12.4%, while maintaining 102 FPS for real-time UAV deployment.
No takes yet. Share an insight, caveat, or question.
xia et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: