MLE-YOLOv11n: Multi-scale layer aggregation and context-aware feature fusion for insulator defect detection in aerial imagery.

Abstract

Power transmission insulators develop defects from sustained high-voltage stress, thermal cycling, and environmental contamination, with undetected degradation progressively leading to self-explosion, flashover, and mechanical failure. Unmanned aerial vehicle (UAV)-based defect detection faces persistent challenges from small object scales, complex background clutter, and multi-scale feature misalignment in lightweight deployable architectures. This paper proposes MLE-YOLOv11n (Multi-scale Layer aggregation and context-aware feature fusion Enhanced YOLOv11n), an enhanced detection framework built upon YOLOv11n that introduces three targeted module substitutions at distinct architectural stages. The Spatial Pyramid Pooling with Efficient Layer Aggregation Network (SPPELAN) replaces the standard Spatial Pyramid Pooling Fast (SPPF) module at the backbone terminus, preserving intermediate pooling representations at each aggregation depth to enrich multi-scale semantic features entering the neck. The Multi-branch Feature Context-Aware (MFCA) attention module replaces equal-weight concatenation at each Path Aggregation Network (PANet) neck fusion node, integrating parallel multi-branch feature enhancement with asymmetric kernels, learnable per-channel adaptive weighting, and cross-region spatial context modeling to improve neck feature discriminability. The Mamba-based Local-Long Attention (MLLA) block replaces the standard convolutional prediction unit at the P4-scale detection head, achieving linear-complexity 𝒪(N) global structural consistency verification through window-partitioned attention with an Irregular Serpentine Scan, suppressing false positives from background structures visually similar to genuine defects. Evaluated on two public benchmarks, MLE-YOLOv11n attains 92.1% mean Average Precision at IoU threshold 0.50 (mAP@50) on the China Power Line Insulator Dataset (CPLID) and 95.3% mAP@50 on the Insulator Defect Image Dataset (IDID), representing gains of 4.4 and 5.2 percentage points over baseline YOLOv11n respectively, with the largest improvements concentrated in rare and structurally complex defect categories. The framework introduces only 10.8% additional parameters (2.87 M total) at 7.2 GFLOPs while maintaining 85 frames per second (FPS) inference speed, demonstrating a competitive accuracy-efficiency trade-off for UAV-based power grid inspection, with inference speed and parameter count suitable for further evaluation on resource-constrained edge platforms in future deployment studies.