Research on interpretable methods for detecting elongated objects in power operations

Miao Li, Yanxia Liu
Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing, PMLR 278:498-510, 2025.

Abstract

Power operations take place in high-risk environments, such as high voltage and strong magnetic fields, making standardized procedures crucial. Employing object detection technology to monitor operational compliance enhances electrical safety. However, the presence of numerous elongated objects and background columnar interferences in power operation datasets significantly affects detection accuracy. To address this issue, we explore model structure improvements from an interpretability perspective. Using Grad-CAM heatmap visualization, we analyze the regions where the model focuses on detection targets. We propose a lightweight convolutional attention mechanism, LCA (Lightweight Convolution Attention), which significantly enhances YOLOv7’s attention to elongated targets while reducing the impact of columnar interference. This improves both the model’s robustness and interpretability. Experimental results show that LCA outperforms classical attention modules such as SE, ECA, and CA, while maintaining a minimal parameter size. Specifically, the mAP of the extremely elongated and challenging sample “operatingbar" increased by 4.4%, and the mAP of the small target “wrongglove" improved by approximately 2%. This makes LCA well-suited for detecting elongated targets in complex power operation environments.

Cite this Paper

BibTeX
@InProceedings{pmlr-v278-li25i, title = {Research on interpretable methods for detecting elongated objects in power operations}, author = {Li, Miao and Liu, Yanxia}, booktitle = {Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing}, pages = {498--510}, year = {2025}, editor = {Zeng, Nianyin and Pachori, Ram Bilas and Wang, Dongshu}, volume = {278}, series = {Proceedings of Machine Learning Research}, month = {25--27 Apr}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v278/main/assets/li25i/li25i.pdf}, url = {https://proceedings.mlr.press/v278/li25i.html}, abstract = { Power operations take place in high-risk environments, such as high voltage and strong magnetic fields, making standardized procedures crucial. Employing object detection technology to monitor operational compliance enhances electrical safety. However, the presence of numerous elongated objects and background columnar interferences in power operation datasets significantly affects detection accuracy. To address this issue, we explore model structure improvements from an interpretability perspective. Using Grad-CAM heatmap visualization, we analyze the regions where the model focuses on detection targets. We propose a lightweight convolutional attention mechanism, LCA (Lightweight Convolution Attention), which significantly enhances YOLOv7’s attention to elongated targets while reducing the impact of columnar interference. This improves both the model’s robustness and interpretability. Experimental results show that LCA outperforms classical attention modules such as SE, ECA, and CA, while maintaining a minimal parameter size. Specifically, the mAP of the extremely elongated and challenging sample “operatingbar" increased by 4.4%, and the mAP of the small target “wrongglove" improved by approximately 2%. This makes LCA well-suited for detecting elongated targets in complex power operation environments. } }
Endnote
%0 Conference Paper %T Research on interpretable methods for detecting elongated objects in power operations %A Miao Li %A Yanxia Liu %B Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing %C Proceedings of Machine Learning Research %D 2025 %E Nianyin Zeng %E Ram Bilas Pachori %E Dongshu Wang %F pmlr-v278-li25i %I PMLR %P 498--510 %U https://proceedings.mlr.press/v278/li25i.html %V 278 %X Power operations take place in high-risk environments, such as high voltage and strong magnetic fields, making standardized procedures crucial. Employing object detection technology to monitor operational compliance enhances electrical safety. However, the presence of numerous elongated objects and background columnar interferences in power operation datasets significantly affects detection accuracy. To address this issue, we explore model structure improvements from an interpretability perspective. Using Grad-CAM heatmap visualization, we analyze the regions where the model focuses on detection targets. We propose a lightweight convolutional attention mechanism, LCA (Lightweight Convolution Attention), which significantly enhances YOLOv7’s attention to elongated targets while reducing the impact of columnar interference. This improves both the model’s robustness and interpretability. Experimental results show that LCA outperforms classical attention modules such as SE, ECA, and CA, while maintaining a minimal parameter size. Specifically, the mAP of the extremely elongated and challenging sample “operatingbar" increased by 4.4%, and the mAP of the small target “wrongglove" improved by approximately 2%. This makes LCA well-suited for detecting elongated targets in complex power operation environments.
APA
Li, M. & Liu, Y.. (2025). Research on interpretable methods for detecting elongated objects in power operations. Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing, in Proceedings of Machine Learning Research 278:498-510 Available from https://proceedings.mlr.press/v278/li25i.html.

Related Material