Assessing the effectiveness of one-stage and two-stage methods for identifying high-voltage power grid equipment in UAV imagery

Unmanned Aerial Vehicle (UAV) imagery, augmented by advanced deep learning architectures, has become an integral approach to the automated inspection and structural monitoring of high-voltage power grids. This study evaluates the practical applicability and speed-accuracy trade-offs of single-stage versus two-stage object detection models for identifying critical power grid components. Specifically, two highly optimized networks were developed and empirically validated: HVE-YOLO11, which uses the latest YOLO11 architecture enhanced with spatial attention mechanisms, and HVE-MASK-R-CNN, which uses a rigorous ResNet-101 Feature Pyramid Network (FPN) backbone. Leveraging a newly curated, diverse dataset of 51,800 augmented images capturing six distinct equipment classes under fluctuating meteorological conditions, the models were evaluated for Mean Average Precision (mAP) and computational throughput measured in Frames Per Second (FPS). Empirical results demonstrate that the single-stage HVE-YOLO11 shatters the traditional speed-accuracy dichotomy, significantly outperforming the two-stage model in both inference velocity (162.9 FPS versus 75.8 FPS in intensive benchmarking) and spatial accuracy (an mAP@0.5 of 0.972 compared to 0.855). These findings provide actionable, highly quantified benchmarks for deploying real-time, AI-driven diagnostic systems on hardware-constrained edge-computing UAV platforms.