Insulator Defect Detection by Fusing Fine-Grained Features and Contextual Information

Abstract

Insulator defect detection is a key task in uncrewed aerial vehicle (UAV) inspection of transmission lines. Defects in real inspection images commonly exhibit small scale, weak texture, blurred boundaries, and strong background interference, leaving existing detectors insufficient in fine-detail feature representation and precise localization. Building upon the end-to-end framework RT-DETR, this paper proposes GLFENet, an insulator defect detection network that integrates global–local context modeling with joint frequency-domain and edge enhancement. Guided by the twin objectives of fine-grained feature preservation and multi-level contextual modeling, GLFENet introduces coordinated improvements in four stages: backbone extraction, context encoding, multi-scale fusion, and bounding-box regression. DCRB enhances shallow local structural representation through dual-branch convolution and multi-scale dilated attention; GLCE combines local-window responses with pooled-global context on high-level features; EGFPN, via its core unit EGFB, fuses spatial-gradient responses with learned frequency-domain features to improve defect distinguishability after multi-scale fusion; and Inner-IoU uses scaled auxiliary boxes to refine bounding-box regression. On the newly constructed FLOWID dataset, GLFENet achieves 63.1% mAP, 93.7% AP50, 72.0% AP75, and 90.0% Recall, improving mAP by 1.5 percentage points and AP75 by 3.6 percentage points over the best baseline; the four modules jointly raise mAP from a baseline of 55.9% to 63.1% and AP75 from 63.2% to 72.0%.