Article Details

Abstract

To address the problems of insufficient illumination, weakened edges, and increased noise in transmission line inspection images under low-light conditions, a hazard detection method based on LLE-DETR is proposed. Built upon RT-DETR, the proposed method incorporates a Low-Light Edge Enhancement Network (LLENet) into the backbone to improve target contours and texture representation. In the hybrid encoder, a Dynamic Adaptive Intra-scale Feature Interaction module (DA-AIFI) is designed to enhance high-level semantic modeling in low-light scenes. In addition, Linear Deformable Convolution (LDConv) is introduced in the cross-scale fusion stage to alleviate the loss of boundary details caused by fixed downsampling. Experimental results show that, on a self-constructed low-light transmission line hazard dataset, the proposed method improves mAP@0.5 and mAP@0.5:0.95 by 3.6% and 2.5%, respectively, compared with the baseline RT-DETR. In addition, the inference speed reaches 53.6 frames per second, demonstrating that the proposed method effectively improves hazard detection performance under low-light conditions while maintaining real-time capability.

Keywords

transmission line; low-light detection; hazard detection; RT-DETR; edge enhancement

References

1. [1] Liu Min, Chen Ming, Wu Minghu, etc GAE-YOLO: Global Perception Enhanced Target Detection Method for External Hidden Hazards in Transmission Lines [J]. Journal of Instrumentation, 2025, 46 (2): 267-278
2. [2] Gao Chao, Wang Zhen, Wu Qiwei, etc Effective external vulnerability detection method for transmission channels based on specific area target detection [J]. Power Grid and Clean Energy, 2025, 41 (1): 44-49, 60
3. [3] Wang Yaru, Feng Lilong, Song Xiaoke, etc TFD-YOLOv8: A Foreign Object Detection Method for Transmission Lines [J]. Journal of Graphic Science, 2024, 45 (5): 901-912
4. [4] Zhao Zhenbing, Zhang Wei, Zhai Yongjie, etc Concept, Research Status, and Prospects of Power Vision Technology [J]. Electric Power Science and Engineering, 2020, 36 (1): 1-8
5. [5] Shi Jintao, Li Zhe, Gu Chaoyue, etc Faster R-CNN power grid foreign object monitoring technology based on sample expansion [J]. Power Grid Technology, 2020, 44 (1): 44-51
6. [6] Zhao Lin, Wang Suzhen, Shao Mingwei, etc Bird's Nest Defect Detection Method for Transmission Lines Based on Improved YOLOv5 [J]. Electronic Measurement Technology, 2023, 46 (3): 157-165
7. [7] Wang Zhenzhou, Sun Dongdong, Wang Jianchao, etc Research on Improved RT-DETR Algorithm for Foreign Object Detection in Transmission Lines [J]. Computer Engineering and Applications, 2026, 62 (2): 116-125
8. [8] Carion N，Massa F，Synnaeve G，et al. End-to-End Object Detection with Transformers [C]//European Conference on Computer Vision，2020：213-229.
9. [9] Zhao Y A，Lyu W Y，Xu S L，et al. DETRs Beat YOLOs on Real-Time Object Detection [PP/OL].arXiv:2304.08069 [cs.CV].https://arxiv.org/abs/2304.08069.
10. [10] Zhang X，Song Y Z，Song T T，et al. LDConv: Linear Deformable Convolution for Improving Convolutional Neural Networks [J/OL].Image and Vision Computing，2024，149：105190.https://doi.org/10.1016/j.imavis.2024.105190.
11. [11] Zhu J C，Chen X L，He K M，et al. Transformers without Normalization [C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)，2025：14901-14911.
12. [12] Liu S L，Li F，Zhang H，et al. DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR [PP/OL].arXiv:2201.12329 [cs.CV].https://arxiv.org/abs/2201.12329.
13. [13] Zhang H，Li F，Liu S L，et al. DINO: DETR With Improved DeNoising Anchor Boxes For End-To-End Object Detection [PP/OL].arXiv:2203.03605 [cs.CV].https://arxiv.org/abs/2203.03605.

Funding

National Natural Science Foundation of China (52074305); Henan Province Science and Technology Research Project (212102210005)

About the authors

Wu Jun (1979-), male, Han ethnicity, from Chizhou, Anhui, professor, doctoral student, research interests include mine informatization and deep learning; Guo Shuaibang (2001-), male, Han ethnicity, from Zhoukou, Henan, master's student, research interests include deep learning and object detection; Leng Qiyuan (2001-), male, Han ethnicity, from Xinxiang, Henan, master's student, research interests include artificial intelligence and deep learning.