Abstract: Deep learning-based foreign object detection techniques for power transmission lines have become a critical enabler for the stable operation of next-generation power grids. To further enhance detection performance for foreign objects on power transmission lines, a hypergraph-based position-aware and global-aware model was proposed, which builds upon Hyper-YOLO’s ability to model and propagate high-order semantic correlations across levels and positions. Firstly, a position-aware multi-branch feature aggregation module (PMFA) was introduced to facilitate precise identification of critical feature representations. Additionally, a global spatial pyramid pooling-fast (Global SPPF) module, equipped with a global receptive field, was designed to enhance boundary-level features and promote the modeling of high-order semantic correlations during feature extraction processing on the backbone. Finally, a lightweight attention-guided weighted downsampling (LAWD) module was proposed to retain critical information as effectively as possible from low-quality input samples. The detection performance of the proposed model was systematically evaluated through ablation studies, comparative analyses, and robustness testing via image augmentation on the collected dataset. Experimental results demonstrate that the proposed model achieves state-of-the-art recognition performance, with a precision of 90.6%, a recall of 87.6%, and a mAP@50 of 93.4%, respectively, and compared with the baseline model, its number of parameters is reduced by 20.5%, and the floating-point computation is reduced by 20.4%, which provides technical support for the detection and maintenance of foreign object on transmission lines.

摘要： 基于深度学习的输电线路异物检测方法已成为确保新一代电网网络能持续稳定运行的重要手段之一。为提高输电线异物检测模型的整体检测性能，利用Hyper-YOLO模型能对特征图中的高阶语义相关性进行跨层和跨位置的建模与传播的特性，提出一种基于超图结构的具有全局视角及位置感知特性的输电线异物检测模型。首先，构建一个具有位置感知特性的多路径特征聚合网络PMFA(Position-aware Multi-branch Feature Aggregation)，以精准地定位需要表征的关键特征。此外，设计具有全局视野的Global SPPF (Global Spatial Pyramid Pooling-Fast)模块，增强骨干网络特征提取的边缘特性以强化检测模型对高阶语义相关性的建模与学习能力。最后，提出基于注意力导向的LAWD(Lightweight Attention-guided Weighted Downsampling)下采样模块，以尽可能地保留低质量数据样本中更多的关键信息。通过一系列的消融实验、对比实验以及数据增强的方式验证了所提出模块在输电线异物检测场景下的有效性及检测性能上的优越性。实验结果表明，所提出模型在构建的输电线异物图像数据集上其检测精度，召回率及mAP@50值分别达到了90.6%、87.6%和93.4%且较之基线模型其参数量降低了20.5%，浮点运算量下降了20.4%，为输电线路场景下异物目标的检测与治理提供了技术支持。