PLFD-YOLO: A Lightweight Multi-Scale Detector for Power Line Defects in UAV Imagery

<title>Abstract</title> <p>Automated visual inspection of power transmission lines via Unmanned Aerial Vehicles (UAVs) provides an essential diagnostic mechanism for grid asset management. Standard single-stage object detectors, however, often struggle with micro-scale defect localization, dense target overlapping, and natural background clutter while maintaining high inference throughput. We present PLFD-YOLO, a lightweight detection model configured specifically for aerial transmission line imagery. The architecture incorporates a four-head Path Aggregation Feature Pyramid Network (PAFPN) featuring a stride-4 P2 branch for high-resolution coordinate preservation, coupled with ADown downsampling layers and Deformable Convolution v2 (DCNv2) blocks to account for non-rigid component deformations. To mitigate background noise, multi-scale feature maps are fused via a learnable channel-weighting module (BiFPNConcat) integrated with direction-aware Coordinate Attention layers. During training, sample assignment is decoupled using bounded CIoU matching to prevent non-zero background similarity from corrupting anchor allocation, while coordinate regression is guided by a size-decay Soft-NWD-CIoU loss function. Evaluated on the PTL-AI Furnas dataset at 1280 × 1280 resolution, PLFD-YOLO obtains 91.57% mAP₅₀ and 64.35% mAP₅₀–₉₅, representing improvements of 2.91 and 4.68 percentage points over YOLO11n, respectively, at an inference throughput of 80.2 FPS on an RTX 5080 workstation, demonstrating strong potential for future embedded UAV deployment.</p>