Object detection of overhead power line components, defects, and foreign objects via multi-source co-occurrence reasoning

Object detection of Overhead Power Line (OPL) components, defects, and foreign objects is of considerable significance for ensuring the safe and stable operation of power systems. To address the prominent intra-class visual appearance variations, multi-scale variations, and significant aspect ratio differences of objects in aerial imagery, this paper proposes MMGY-Net (Multi-source integration, Multi-scale representation, and Geometry-aware You Only Look Once (YOLO)-based Network). First, a Multi-source Co-occurrence Reasoning Graph Convolutional Network (MSCR-GCN) is designed to fuse multi-scale global visual features, category instance-level features, category-level semantic embeddings, and category co-occurrence priors, thereby achieving joint reasoning over multi-source information and enhancing the representational and discriminative capabilities of the model. Second, a Fine-grained Compensatory Cross-scale Fusion Network (FCCFN) is designed to enhance the model’s capability to exploit and fuse multi-scale features. Finally, a Geometry-aware Intersection over Union (GAIoU) loss function is developed, which decouples the geometric constraints along the width and height dimensions of bounding boxes and jointly optimizes the regression process with an auxiliary bounding box, enabling the model to more accurately delineate the boundaries of objects with pronounced geometric shape variations. On the self-constructed 10 kilovolt (kV) OPL Components, Defects, and Foreign Objects (10-kV-OPL-CDF) dataset, MMGY-Net achieves 88.0% mean Average Precision (mAP)@0.5 and 59.9% mAP@0.5:0.95, with 13.4 Frames Per Second (FPS) for engine-only inference on the Jetson Nano Developer Kit. Furthermore, experiments on public datasets demonstrate competitive performance across diverse OPL scenarios. MMGY-Net provides a detection method that balances accuracy and efficiency for OPL inspection.