HVLO-YOLO: An Ultra-Lightweight Detection Model for High-voltage Line Obstacles


Weichao Pan (Shandong Jianzhu University), Xu Wang (Shandong Jianzhu University), Chengze Lv (Shandong Jianzhu University), Zicheng Lin (Shandong Jianzhu University), Gongrui Wang (Shandong Jianzhu University), Xuening Zhang (Harbin Institute of Technology), Yi Sun (University of Ulster), Xingbo Liu (Shandong Jianzhu University)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)

Abstract

With the expansion of high-voltage power grids and the increase of environmental complexity, obstacle detection on high-voltage lines has become an important task to ensure the safety of power systems. Traditional methods rely on manual feature extraction, which is difficult to deal with complex environments. Although deep learning methods improve the detection accuracy, the demand for computing resources is too high to meet the requirements of real-time and lightweight. To this end, this paper proposed an ultra-lightweight \textbf{H}igh-\textbf{V}oltage \textbf{L}ine \textbf{O}bstacle detection model \textbf{HVLO-YOLO}. To achieve a better balance between detection accuracy and computational cost, three specialized lightweight modules are introduced: (1) a CSP-Partial Convolution with FourGroup module that enhances feature extraction efficiency by selectively applying partial convolution and multi-branch group strategies; (2) a Partial Convolution DownSampler module that preserves critical information during spatial resolution reduction through a dual-branch design combining max-pooling and partial convolution; and (3) a Partial Convolution Detection Head module that focuses computational resources on key feature regions through selective lightweight aggregation. These modules collaboratively reduce computational burden, minimize parameter count, and enhance obstacle detection accuracy under complex environments. Extensive experiments conducted on two benchmark datasets demonstrate that HVLO-YOLO achieves competitive detection accuracy while significantly reducing model complexity compared to state-of-the-art models. Our code is publicly available at \url{https://github.com/JEFfersusu/HVLO-YOLO}.

Citation

@inproceedings{Pan_2025_BMVC,
author    = {Weichao Pan and Xu Wang and Chengze Lv and Zicheng Lin and Gongrui Wang and Xuening Zhang and Yi Sun and Xingbo Liu},
title     = {HVLO-YOLO: An Ultra-Lightweight Detection Model for High-voltage Line Obstacles},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_405/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection