电子测量与仪器学报

王凌云,李婷宜,李阳,万旭东,童华敏.基于 FEF-DeepLabV3+的电力金具锈蚀分割方法[J].电子测量与仪器学报,2023,37(7):166-176

基于 FEF-DeepLabV3+的电力金具锈蚀分割方法

Segmentation method of power armor clamp corrosion based on FEF-DeepLabV3+

DOI：

英文关键词:deep learning overhead transmission line inspection image semantic segmentation defect detection pixel classification

基金项目:国家自然科学基金(51907104)项目资助

作者	单位
王凌云	1. 三峡大学电气与新能源学院
李婷宜	1. 三峡大学电气与新能源学院
李阳	2. 国网宜昌供电公司
万旭东	1. 三峡大学电气与新能源学院
童华敏	2. 国网宜昌供电公司

Author	Institution
Wang Lingyun	1. College of Electrical Engineering and New Energy, China Three Gorges University
Li Tingyi	1. College of Electrical Engineering and New Energy, China Three Gorges University
Li Yang	2. State Grid Yichang Power Supply Company
Wan Xudong	1. College of Electrical Engineering and New Energy, China Three Gorges University
Tong Huamin	2. State Grid Yichang Power Supply Company

摘要点击次数: 892

全文下载次数: 893

中文摘要:

金具锈蚀在输电线路航拍图像中细节丰富且分布不规律,为克服分割检测过程中局部信息丢失、精度低和速度慢等问题,提出基于 DeepLabV3+的金具锈蚀语义分割模型。替换其主干网络为轻量化改进 MobileNetV3 网络加快运算速度,并提出自适应特征金字塔( adaptive feature pyramid,AFP) 结构融合多尺度。结合 FRN 层提出特征融合空洞空间金字塔池化( feature fusion atrous spatial pyramid pooling,FEF-ASPP)结构,能够在加强像素间联系的同时不降低分辨率。最后优化损失函数,提高算子的有效性。实验表明,mIoU 和 mPA 分别达到了 87. 15%、96. 64%,相较于原模型提高了 3. 09%、4. 29%。参数量仅为原模型的 48%,推理时间仅为 15. 94 ms,降低了对设备算力的要求,实现高效高精度、轻量化的输电设备锈蚀缺陷分割检测。

英文摘要:

The proportion of armor clamp rust in aerial images of power transmission lines is rich in details and irregularly distributed. To overcome problems such as local information loss, low accuracy, and slow speed in the segmentation detection process, a DeepLabV3+- based semantic segmentation model for armor clamp rust is proposed. The backbone network is replaced with a lightweight improved MobileNetV3 network to speed up computation, and an adaptive feature pyramid (AFP) structure is proposed to merge multiple scales. A feature fusion atrous spatial pyramid pooling ( FEF-ASPP) structure is proposed, combined with the FRN layer to strengthen pixel relationships without reducing resolution. Finally, the loss function is optimized to improve the effectiveness of the operator. Experiments show that the mIoU and mPA reach 87. 15% and 96. 64%, respectively, which is an improvement of 3. 09% and 4. 29% compared to the original model. The parameter quantity is only 48% of the original model, and the inference time is only 15. 94 ms, reducing the requirement for device computing power and achieving high-efficiency, high-precision, and lightweight segmentation detection of armor clamp rust in power transmission equipment.

查看全文查看/发表评论下载PDF阅读器