Topology-guided lightweight multi-scale network for component recognition and defect detection in overhead distribution lines

Introduction Accurate defect detection in overhead distribution-line imagery requires both local visual evidence and structural information about the associated electrical components. This study proposes TG-LMNet, a lightweight multi-scale network for component recognition and defect localization. Methods TG-LMNet incorporates predicted component locations, categories, and confidence scores into defect detection through spatial compatibility fields. An anisotropic backbone preserves directional features, weighted multi-scale fusion integrates local and contextual information, and a topology-consistency objective constrains component-defect relationships. Local boundary supervision is additionally applied when mask annotations are available. Results On the synthetic SD-ODLD benchmark containing 7,200 images, nine component classes, and eight defect classes, TG-LMNet achieves 85.6% small-defect recall and a 5.8% image-level false-alarm rate, with 9.6 million parameters and 23.8 GFLOPs. On a label-aligned subset of real UAV inspection imagery, the model obtains 72.8% mAP@0.5 under zero-shot transfer and 81.3% after few-shot adaptation. The reported overall host-component association accuracy is 92.4%. Discussion The results indicate the potential of component-conditioned representations for defect detection while revealing a measurable gap between synthetic and real inspection imagery. Further evaluation is needed to establish generalization across inspection conditions.