基于U-net和YOLOv4的绝缘子图像分割与缺陷检测

唐小煜; 黄进波; 冯洁文; 陈锡和

doi:10.6054/j.jscnun.2020088

基于U-net和YOLOv4的绝缘子图像分割与缺陷检测

doi: 10.6054/j.jscnun.2020088

广东省量子调控工程与材料重点实验室//广东省光电检测仪器工程技术研究中心// 物理国家级实验教学示范中心//华南师范大学物理与电信工程学院，广州 510006

基金项目:

国家自然科学基金项目 61371176

广州市高校创新创业教育项目 2019HD206

详细信息

通讯作者:
唐小煜，讲师，Email:tangxy@scnu.edu.cn

中图分类号: P407.8;TP75
计量
- 文章访问数: 1053
- HTML全文浏览量: 366
- PDF下载量: 105
- 被引次数: 0
出版历程
- 收稿日期: 2020-05-22
- 刊出日期: 2020-12-25

Image Segmentation and Defect Detection of Insulators Based on U-net and YOLOv4

Guangdong Provincial Key Laboratory of Quantum Engineering and Quantum Materials//Guangdong Provincial Engineering Technology Research Center for Optoelectronic Instrument//National Demonstration Center for Experimental Physics Education//School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou 510006, China

摘要

摘要: 输电线路上绝缘子的完整性直接影响了输电的安全与可靠性.采用深度学习方法，对绝缘子图像识别提取和缺陷检测问题进行了研究.首先基于优化的U-net模型获取绝缘子区域掩模图像，实现对绝缘子串语义分割；然后基于YOLOv4模型获取缺陷绝缘子的位置，实现对自爆绝缘子目标的检测.为充分利用高分辨率图像的像素信息，提出“切分-识别-合成”的检测思路，精确分割出绝缘子以及判断并获取缺陷区域；最后设计了多组实验并进行对比，验证了模型的有效性.采用优化的U-net模型分割绝缘子的Dice系数达0.92；采用YOLOv4模型检测自爆绝缘子的识别精度达0.96，平均重叠度IOU达0.88.研究结果对实现电力系统运维的智能化具有较高的应用价值.
- 绝缘子 /
- 语义分割 /
- 目标检测 /
- U-net /
- YOLOv4
Abstract: The integrity of the insulator on transmission lines directly affects the safety and reliability of power transmission. Deep learning methods are used to explore the problems of insulator image extraction and defect detection. First, the mask image of the insulator region is obtained with the optimized U-net model so as to implement the semantic segmentation of the insulator. Then the YOLOv4 model is built to obtain the position of the defective insulator and achieve the detection of the self-explosive insulator. Aiming at making full use of the pixel information of high-resolution images, the "segmentation-recognition-synthesis" model is proposed to segment the insulator and obtain the defect area precisely. Multiple groups of experimental comparison are designed to verify the effectiveness of the model. The Dice coefficient of insulator segmentation based on the optimized U-net model is up to 0.92; the accuracy of the self-explosive insulator identification based on the YOLOv4 model is up to 0.96, and the average degree of overlapping IOU reaches 0.88. The research results have high practical value for the realization of intelligent power system operation and maintenance.
- insulator /
- semantic segmentation /
- object detection /
- U-net /
- YOLOv4

HTML全文

图 1 U-net模型框架

Figure 1. The U-net model framework

下载: 全尺寸图片幻灯片

图 2 绝缘子串珠分割的流程图

Figure 2. The flow chart of semantic segmentation of the insulator

下载: 全尺寸图片幻灯片

图 3 不同数据集的分割结果

Figure 3. The segmentation results of different data sets

下载: 全尺寸图片幻灯片

图 4 YOLOv4框架

Figure 4. The YOLOv4 framework

下载: 全尺寸图片幻灯片

图 5 训练过程损失值和均值平均精度曲线

Figure 5. The loss and map curve of the training process

下载: 全尺寸图片幻灯片

图 6 绝缘子缺陷检测流程图

Figure 6. The flow chart of insulator defect detection

下载: 全尺寸图片幻灯片

图 7 滑动窗口切分示意图

Figure 7. The diagram of sliding window segmentation

下载: 全尺寸图片幻灯片

图 8 训练损失曲线与测试Dice系数曲线

Figure 8. The training loss curve and test Dice curve

下载: 全尺寸图片幻灯片

图 9 FCN模型与优化后U-net模型的分割效果对比

Figure 9. The comparison of segmentation results between the FCN model and the optimized U-net model

下载: 全尺寸图片幻灯片

图 10 优化后测试的效果图

Figure 10. The test results after optimization

下载: 全尺寸图片幻灯片

表 1 FCN和优化后U-net的效果对比

Table 1. The comparison of results between the FCN model and the optimized U-net model

模型	图像数	精确率/%	召回率/%	Dice/%	平均处理时间/ms
FCN	126	91.83	81.16	85.56	202
优化后U-net	126	94.34	89.05	91.31	191

下载: 导出CSV

表 2 检测优化前后的性能

Table 2. The performance before and after detection optimization

模型	图像数	平均处理时间/ms	识别精确率/%	平均IOU
优化的YOLOv4	304	53	0.96	0.88
YOLOv4	304	53	0.88	0.76
Faster-RCNN	304	605	0.93	0.83

下载: 导出CSV

参考文献(13)

[1]	黄新波, 刘新慧, 张烨, 等.基于红蓝色差和改进K-means算法的航拍绝缘子分类识别方法[J].高电压技术, 2018, 44(5):1528-1534. http://www.cnki.com.cn/Article/CJFDTotal-GDYJ201805018.htm HUANG X B, LIU X H, ZHANG Y, et al. Classification recognition method of insulator in aerial image based on the red-blue difference and developed K-means algorithm[J]. High Voltage Engineering, 2018, 44(5):1528-1534. http://www.cnki.com.cn/Article/CJFDTotal-GDYJ201805018.htm
[2]	刘永权, 肖德军.基于支持向量机的绝缘子图像分割[J].科技与企业, 2015(22):250-250. http://www.cqvip.com/main/zcps.aspx?c=1&id=666629088 LIU Y Q, XIAO D J. Insulator image segmentation based on support vector machine[J]. Technology and Enterprise, 2015(22):250-250. http://www.cqvip.com/main/zcps.aspx?c=1&id=666629088
[3]	梁耀, 黎双文, 刘鑫磊, 等.复杂背景下目标树叶自动分割的Grabcut算法[J].华南师范大学学报(自然科学版), 2018, 50(6):112-118. doi: 10.6054/j.jscnun.2018126 LIANG Y, NI S W, LIU X L, et al. The grabcut algorithm for the automatic segmentation of target leaves under the complex background[J]. Journal of South China Normal University (Natural Science Edition), 2018, 50(6):112-118. doi: 10.6054/j.jscnun.2018126
[4]	高金峰, 吕易航.航拍图像中绝缘子串的识别与分割方法研究[J].郑州大学学报(理学版), 2019, 51(4):16-22. http://www.zhangqiaokeyan.com/academic-journal-cn_journal-zhengzhou-university-natural-science-edition_thesis/0201273466894.html GAO J F, LV Y H. Research on insulator string detection, segmentation and self-explosion fault identification method in aerial images[J]. Journal of Zhengzhou University(Science Edition), 2019, 51(4):16-22. http://www.zhangqiaokeyan.com/academic-journal-cn_journal-zhengzhou-university-natural-science-edition_thesis/0201273466894.html
[5]	REN S, HE K, GIRSHICK R, et al. Faster R-CNN:towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6):1137-1149. doi: 10.1109/TPAMI.2016.2577031
[6]	陈文浩, 姚利娜, 李丰哲.无人机电网巡检中的绝缘子缺陷检测与定位[J].计算机应用, 2019, 39(S1):210-214. http://www.cnki.com.cn/Article/CJFDTotal-JSJY2019S1044.htm CHEN W H, YAO L N, LI F Z. Insulator defect detection and location in UAV grid inspection[J]. Journal of Computer Applications, 2019, 39(S1):210-214. http://www.cnki.com.cn/Article/CJFDTotal-JSJY2019S1044.htm
[7]	王梦.基于绝缘子图像的缺陷检测方法研究[D].武汉: 华中科技大学, 2019. WANG M. Study on the method of fault detection based on insulator images[D]. Wuhan: Huazhong University of Science and Technology, 2019.
[8]	RONNEBERGER O, FISCHER P, BROX T. U-net: convolutional networks for biomedical image segmentation[C]//International Conference on Medical image computing and computer-assisted intervention. Berlin: Springer, 2015: 234-241.
[9]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. New York: IEEE, 2016: 779-788.
[10]	BOCHKOVSKIY A, WANG C Y, LIAO H Y. YOLOv4: optimal speed and accuracy of object detection[EB/OL]. https://arxiv.org/abs/2004.10934.
[11]	WANG C Y, LIAO H Y, WU Y H, et al. CSPNet: a new backbone that can enhance learning capability of cnn[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. New York: IEEE, 2020: 390-391.
[12]	HE K, ZHANG X, REN S, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE transactions on pattern analysis and machine intelligence, 2015, 37(9):1904-1916.
[13]	LIU S, QI L, QIN H, et al. Path aggregation network for instance segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. New York: IEEE, 2018: 8759-8768.