基于OpenCV+SSD深度学习模型的变电站压板状态智能识别

王伟; 张彦龙; 翟登辉; 刘力卿; 许丹; 张旭

王伟,张彦龙,翟登辉,刘力卿,许丹,张旭.基于OpenCV+SSD深度学习模型的变电站压板状态智能识别[J].电测与仪表,2022,59(1):106-112.
WANG Wei,ZHANG Yanlong,ZHAI Denghui,Liu Liqing,Xudan,ZHANG Xu.Intelligent identification of substation platen state based onOpenCV + SSD deep learning model[J].Electrical Measurement & Instrumentation,2022,59(1):106-112.

基于OpenCV+SSD深度学习模型的变电站压板状态智能识别

Intelligent identification of substation platen state based onOpenCV + SSD deep learning model

DOI：10.19753/j.issn1001-1390.2022.01.014

中文关键词: 压板状态识别 OpenCV 透视变换深度学习 SSD目标检测。

英文关键词:Pressure plate state recognition OpenCV Perspective transformation deep learning SSD object detection

基金项目:国家电网公司总部科技项目资助(5206/2018-19002A) “智能决策关键技术研究及应用”

作者中文名	作者英文名	单位
王伟	WANG Wei	国网天津市电力公司电力科学研究院
张彦龙	ZHANG Yanlong	许继电气股份有限公司
翟登辉	ZHAI Denghui	许继电气股份有限公司
刘力卿	Liu Liqing	国网天津市电力公司电力科学研究院
许丹	Xudan	许继电气股份有限公司
张旭	ZHANG Xu	许继电气股份有限公司

摘要点击次数: 2685

中文摘要:

目前,变电站保护硬压板信息管理处于完全依赖于人工巡检的状态。伴随着人工智能上升为国家战略,深度学习技术进一步发展,提出一种基于OpenCV+SSD深度学习模型的压板状态识别方法。首先在图像二值化并高斯滤波基础上,基于霍夫直线检测算法进行保护屏柜角点检测,并通过透视变换实现压板图像矫正,从而避免拍摄图像的角度变化对识别结果的影响；其次,利用模板匹配的方法对压板图像进行分割,并建立不同压板与其功能间的映射关系,提升压板状态检测准确度；然后基于TensorFlow深度学习框架搭建SSD目标检测模型,通过不断调参并采用正则化处理提高训练模型的准确度和泛化能力。经测试验证,该方法目标检测精确率和召回率均大于0.95,相对于传统opencv及Hog+SVM的处理方法,压板状态识别效果有明显提高。

英文摘要:

At present, the information management of substation protection hard plate is totally dependent on manual inspection. With the rise of AI to national strategy and the further development of in-depth learning technology, a plate state recognition method based on in-depth learning OpenCV+SSD network model is proposed. Firstly, on the basis of image binarization, corner detection of protective screen cabinet is carried out based on Hough linear detection algorithm. Through perspective transformation, the plate image is corrected, so as to avoid the influence of the angle change of the photographed image on the recognition results. Secondly, the template matching method is used to segment the image of the press plate, and the mapping relationship between different press plates and their functions is established to improve the accuracy of the state detection of the press plate. Then, a target detection model of SSD is built based on the TensorFlow deep learning framework, and the accuracy and generalization ability of the training model are improved by adjusting the parameters continuously and using regularization processing. The test results show that the target detection accuracy and recall rate of this method are both greater than 0.95. Compared with the traditional opencv and Hog + SVM processing methods, the recognition effect of press plate state has been significantly improved.

查看全文查看/发表评论下载PDF阅读器

关闭