AE-MCDD: Attention-enhanced multiple component defects detection for UAV-assisted powerline inspection

Jiehao Li, Manjia Liu, Haitao Peng, Longlong Liu, Xiaomin Zheng, Chen Yi, Guozi Liu, Jieyu Zhou, Feng Lyu

Abstract: Transmission lines are critical infrastructure for power delivery, yet their exposure to harsh environmental conditions accelerates component degradation, leading to defects that threaten grid reliability. While UAV-assisted inspections enable efficient defect identification, existing automated detection models face persistent challenges including severe class imbalance and complex environmental interference. To address these challenges, we first analyze a real-world 5,300-image aerial dataset, demonstrating severe class imbalance and diverse environmental noise in detail. We then propose a three-pronged solution: (1) a data augmentation pipeline integrating random occlusion and mirroring to enhance rare defect samples; (2) the Attention-Enhanced Multiple Component Defects Detection (AE-MCDD) model, combining HgNetV2 for local feature extraction, a Hybrid Attention Transformer (HAT) module for global context modeling, and C2F modules with skip connections for multi-scale feature fusion; and (3) a focal-loss-optimized multi-task loss function to handle class imbalance. Extensive experiments on our real-world dataset demonstrate that the proposed AE-MCDD model achieves a 0.719 \(m\text {AP}_{50}\), outperforming baseline methods in both common and rare defect detection.

External IDs:dblp:journals/ppna/LiLPLZYLZL25