A Scene Knowledge Integrating Network for Transmission Line Multi-Fitting Detection
Author Information
Author(s): Chen Xinhang, Xu Xinsheng, Xu Jing, Zheng Wenjie, Wang Qianming
Primary Institution: College of Quality & Standardization, China Jiliang University
Hypothesis
The study proposes a Scene Knowledge Integrating Network (SKIN) to improve multi-fitting detection in transmission lines by addressing severe occlusion and tiny-scale object problems.
Conclusion
The proposed SKIN significantly enhances detection performance for occluded and tiny-scale fittings, achieving a 4.8% increase in mAP compared to the baseline model.
Supporting Evidence
- The SKIN model improved detection accuracy for tiny-scale fittings by 11.5%.
- The SKIN model improved detection accuracy for severely occluded fittings by 9.9%.
- The introduction of scene knowledge significantly enhanced the model's performance.
- The model was tested on a dataset of 1330 images containing 16,358 fitting objects.
Takeaway
This study created a smart system to help find small and hidden parts of power lines using special knowledge about how these parts are usually arranged.
Methodology
The study used a deep learning approach with a Scene Knowledge Integrating Network that includes a scene filter module and a scene structure information module to improve detection accuracy.
Limitations
The model still faces challenges with severe occlusion and extreme scale variations.
Statistical Information
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website