GMS ‐ RT ‐ DETR : A Geometry‐Aware and Lightweight Model for Real‐Time Insulator Defect Detection in UAV Power Line Inspection

ABSTRACT To address the critical challenges of complex background interference, subtle micro‐crack detection, and the geometric modeling of slender targets with extreme aspect ratios in UAV power line inspection, this paper proposes GMS‐RT‐DETR (Geometric & Multi‐Scale RT‐DETR), a geometry‐aware and lightweight real‐time detection model. To reduce computational complexity for edge deployment without sacrificing feature representation, the StarNet lightweight architecture is adopted to replace the original ResNet‐18 backbone. By leveraging star operations (element‐wise multiplication), it achieves high‐dimensional nonlinear feature mapping within a compact dimensional space. Furthermore, to address the scale variance between macroscopic insulator contours and microscopic local defects, a Spatial Pyramid Pooling‐Fast (SPPF) module is embedded at the terminal stage of the backbone, utilizing serial multi‐kernel pooling to expand the effective receptive field. Within the subsequent hybrid encoder, a Split‐Directional Aware Attention Module (DAAM) is designed to capture geometric features of slender targets and suppress background noise. This module employs a channel decoupling strategy and asymmetric convolutions to accurately capture linear structures and directional cracks. To further improve bounding box regression for targets with extreme aspect ratios, a Direction‐Aware Intersection over Union (DA‐IoU) loss function is proposed, introducing an aspect‐ratio‐guided dynamic weighting mechanism to adaptively penalize localization errors along the shorter side. Experimental results demonstrate that GMS‐RT‐DETR achieves 94.3% mAP50 and 95.1% precision on the insulator defect dataset, representing absolute improvements of 4.2% and 4.0% over the baseline, respectively. In terms of model efficiency, the parameter count is reduced by 40.7% (from 18.9 to 11.2 M), the computational cost is lowered to 32.5 GFLOPs, and the inference speed reaches 79.4 FPS. These metrics satisfy the real‐time and edge‐deployment requirements of autonomous UAV power line inspection.