集成学习和条件随机场的影像招弧角识别  

Recognition of arcing horns from images using ensemble learning and conditional random field

作　　者：宁淼福[1] 孙阔腾 蔡玮辰 唐鑫 夏立杰 NING Miaofu;SUN Kuoteng;CAI Weichen;TANG Xin;XIA Lijie(Liuzhou Bureau,CSGEHV Power Transmission Company,Liuzhou,Guangxi545006,China)

机构地区：[1]中国南方电网有限责任公司超高压输电公司柳州局,广西柳州545006

出　　处：《测绘科学》2021年第6期114-120,共7页Science of Surveying and Mapping

摘　　要：无人机巡检是目前电力部门主推的一种巡检方式,招弧角是一种重要的电网设备、但其呈现细长的几何形状特征,其测量需要优于1 cm空间分辨率的影像。为了从无人机获取的高分辨率影像上提取招弧角,该文提出了基于随机森林、集成学习、全连接条件随机场的无人机影像分类和招弧角提取方法。首先,提取了影像的12个光谱和纹理特征。接着,建立训练样本库,训练了多个独立的随机森林分类器、并形成随机森林集成模型进行影像分类。最后,利用全连接条件随机场优化分类结果。该文采用5000张无人机影像进行了实验。实验表明,该文提出方法的整体分类精度达到85.5%,招弧角识别的正确率为98.3%、完整率为74.3%,表明该方法具有潜在的工程应用价值。Currently, unmanned aerial vehicle(UAV)-based patrol inspection is one of main ways to patrol inspection in the power sectors, but the geometric measurement of some powerline equipment, such as the arcing horns, need images with the spatial resolution being finer than 1 cm, because an arcing horn is very thin and long. In order to extract arcing horns from the high resolution images obtained by UAV photogrammetry technology, a method for UAV image classification and arcing horns extraction using random forest, ensemble learning and fully connected conditional random field was proposed in this paper. First, twelve features about spectral property and texture were extracted. Then, a training sample database was established to train some independent random forest classifiers, and a random forest ensemble model was formed for image classification. Finally, the classification results were optimized by using fully connected conditional random field. There were 5 000 UAV images being used in the experiment. Experiments showed that the classification accuracy of the proposed method was 85.5%,the accuracy and completeness of arcing horns recognition was 98.3% and 74.3% respectively, which showed that the method had potential engineering application value.

关 键 词：无人机 电力线路巡检 高分辨率遥感影像 随机森林 集成学习 条件随机场 

分 类 号：P237[天文地球—摄影测量与遥感]