Deep Learning-Based Object Detection for Power Line Components: Recent Progress, Current Challenges, and Future Perspectives

Abstract

Unmanned aerial vehicle (UAV)-based inspection is increasingly replacing manual patrols of high-voltage transmission lines, which requires reliable detection of power line components (PLCs). This task remains challenging because the PLCs are small, visually similar, and densely distributed. Furthermore, existing review papers generally address transmission-line inspection broadly, with limited systematic comparison of object detection methodologies under a common evidence-based standard. Hence, this paper presents a systematic survey of deep learning-based PLC detection methods published between 2017 and 2026, indexed in IEEE Xplore, Scopus, Web of Science, and arXiv. A total of 26 approaches are examined across three object detection methodologies: one-stage You Only Look Once, two-stage Faster Region-based Convolutional Neural Network, and transformer-based Detection Transformer, together with six public datasets and seven embedded platforms. To improve cross-study comparability, the reported results are analyzed according to their respective dataset, evaluation metrics, and hardware platform. Across the surveyed studies, the reported ranges of mean average precision at 0.5 intersection over union are from 63.8% to 98.3%. In detecting single-class insulators, the reported performance has largely exceeded 90%, whereas in detecting multi-class cases on datasets such as InsPLAD and STN-PLAD, the performance remains limited due to long-tailed class distributions and reduced recall for rare PLC categories. The reported throughput varies by more than fourfold across different hardware platforms and precision settings, highlighting general weaknesses in real-time performance claims. Finally, several key reproducibility gaps are identified, followed by seven future research directions toward a more reliable and efficient PLC detection for UAV-based utility inspection systems.