摘要
目标检测和识别已经在输电线路巡检中被广泛采用;然而由于宽视场图像数据量大,小目标相对宽视场较小,分辨率低,现有的图像金字塔、特征金字塔和多异构特征融合等方法虽能准确地检测大目标,但小目标的检测精度低,处理非常耗时,因而快速、准确地检测宽视场图像中小目标仍是一个挑战;提出一个两个Faster-RCNs级联的上下文宽视场小目标检测卷积网络,首先,针对降分辨率的宽视场图像,利用一个Faster R-CNN来检测目标的上下文区域,然后,针对上下文区域对应的高分辨率原始图像,利用Faster R-CNN来检测来小目标;用航拍输电线路图像数据集进行了多尺度目标的检测试验,试验结果表明,文章提出的目标检测方法达到了88%的检测精度,检测精度明显优于单级Faster R-CNN检测方法。
Object detection and recognition has been widely applied to power transmission line inspection.Existing methods,such as multi-scale image pyramid,multi-scale feature pyramid and multiple heterogeneous feature fusion,etc.can detect small objects accurately,but usually require heavy computational burden,thus fast and precise target detection in wide-view-field images is still challenging due to large amount of image data and low resolution of small targets.In this paper,we propose a two-stage context convolutional network for small target detection in wide-view-field images,which consists of two cascaded Faster R-CNNs,thefirst Faster R-CNN is used to locate context regions in a low resolution image,and another Faster R-CNN to detect small targets in high-resolution images of detected context regions.We test the proposed method is test on our data sets captured by unmanned aircraft,experimental results show that the proposed method could lead to 88.0%accuracy for small target detection and is higher than that of the one-stage Faster R-CNN.
作者
王海涛
姜文东
程远
严碧武
张宗峰
李涛
张森海
Wang Haitao;Jiang Wendong;Cheng Yuan;Yan Biwu;Zhang Zongfeng;Li Tao;Zhang Senhai(Wuhan NARI Limited Liability Company,State Grid Electric Power Research Institute,Wuhan 430074,China;NARI Group Corporation Ltd. Nanjing 211106,China;State Grid Zhejiang Electric Power Co.,Ltd.,Hangzhou 310007,China;Rizhao Power Supply Company,State Grid Shandong Electric Power Company,Rizhao 276826,China;State Grid Jiaxing Power Supply Company,Jiaxing 314003,China)
出处
《计算机测量与控制》
2019年第6期199-204,共6页
Computer Measurement &Control
基金
国家电网公司科技项目(521104180025)
关键词
小目标检测
无人机图像
输电线路巡检
small object detection
UAV images
power transmission line
