基于语义分割的接触网开口销状态检测

The state detection of split pin in overhead contact system based on semantic segmentation

针对铁路智能巡检系统中开口销缺陷样本不足的问题,通过改变传统接触网部件状态检测思路,提出了一种基于图像语义分割的开口销状态检测方法.将开口销语义信息分为头部信息与尾部信息并进行多边形标注,训练DeepLabv3+模型,提取开口销的信息,分析开口销头部与尾部连通域及其之间的关系,从而判断开口销状态.使用了语义分割的方法,在训练过程中只使用正常的开口销样本,无须专门搜集或制作开口销缺陷样本.验证算法的检测精度,取开口销正常状态、缺失状态、松脱状态和非开口销区域的样本数分别为1 000,20,50和1 000,识别率分别达到95.3%,100.0%,84.0%和98.7%.

Aiming at the problem of insufficient sample of opening pin defect in railway intelligent inspection system,a method for detecting the state of split pin based on image semantics segmentation was proposed,which changes the traditional method of state detection of overhead contact system(OCS).Semantic information of the split pin is divided into head and tail for polygon labeling,DeepLabv3+ model is trained,and the information of the split pin is extracted.Analyzing the connection area between the head and tail of the split pin and its relationship,so as to judge the state of the split pin.With the method of semantics segmentation,only positive samples of split pins are used in the training process,so there is no need to collect or fabricate negative samples of split pins.In order to verify the accuracy of the algorithm,1 000,20,50 and 1 000 samples were taken from the normal, missing,loosed and non-split areas,and the detection accuracy was 95.3%,100.0%,84.0% and 98.7% respectively.