Fog-Adaptive-YOLO: A lightweight model for insulator defect detection

Abstract

Insulator defect detection under foggy conditions suffers from complex backgrounds, small targets, weak features, and severe weather interference, remaining a challenging task for UAV-based inspection. To address these issues, this paper proposes Fog-Adaptive-YOLO, a lightweight fog-adaptive detection network. The FogEnhance module suppresses fog noise and enhances weak defect features; the C3MSGR and C2fMSGR modules optimize lightweight multi-scale feature extraction and aggregation. Experimental results show that on the self-constructed InsDef-Fog dataset, the proposed model achieves 65.4% mAP50 with only 2.74M parameters. It obtains 60.3% mAP50 on the public IDID_FOG dataset and 80.2% mAP50 on the real-world WM-FOG dataset. The model also maintains stable precision on the cross-scene RTTS foggy dataset. These results demonstrate that Fog-Adaptive-YOLO achieves a favorable balance between detection accuracy and lightweight efficiency, well-suited for practical foggy insulator defect detection tasks.