可变形网络与迁移学习相结合的电力塔遥感影像目标检测法  

Power tower detection in remote sensing imagery based on deformable network and transfer learning

作　　者：郑鑫 潘斌[1] 张健 ZHENG Xin;PAN Bin;ZHANG Jian(School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China)

机构地区：[1]武汉大学遥感信息工程学院,湖北武汉430079

出　　处：《测绘学报》2020年第8期1042-1050,共9页Acta Geodaetica et Cartographica Sinica

摘　　要：电力塔是电力基础设施的重要组成部分,对其进行检测是必不可少的工作。针对当前遥感影像电力塔检测算法精度低,效果差的问题,本文基于可变形网络和迁移学习对Faster R-CNN进行改进,提出一种基于遥感影像的电力塔检测框架。该框架主要分为两个部分:①特征提取子网络,即利用可变形网络模型改进卷积层,来提高模型对于电力塔几何形变的特征提取能力;②目标检测子网络,即通过模型迁移,将由特征提取子网络训练获得的模型参数迁移至此子网络,由RPN网络和可变形区域池化结合非极大值抑制(NMS)精确获取电力塔位置,利用Fine-tuning技术快速训练此子网络,最终实现高精度的遥感影像电力塔检测。本文算法在测试集中对电力塔检测结果为AP0.50.8861,AP0.60.8396,ACC 0.8948,与SSD、YOLOv3、Faster R-CNN等相比,各检测指标至少高0.2。由对比试验可以看出,该框架对电力塔遥感影像可以实现较高精度检测,表明该方法在电力塔检测上拥有较大应用潜力。Power towers are important parts of power infrastructure,and it is indispensable to detect them.In view of the low precision and poor result of detection algorithms for power towers in remote sensing imagery,this study improves Faster R-CNN based on deformable network and transfer learning.And then we propose a new detection framework for power tower in remote sensing imagery.The framework includes a feature extraction sub-network and an object detection sub-network.The feature extraction sub-network uses deformable network model,which reconstructs the convolutional layer,to improve the model’s feature extraction ability of the power towers with geometric deformation.The model parameters obtained from the feature extraction sub-network training are transferred to object detection sub-network,which accurately obtains position of power towers through RPN network,deformable area pooling and nms algorithms.Finally,the object detection sub-network is finely tuned and achieve high-precision detection for power towers in remote sensing image.The results show that in the test datasets AP0.5,AP0.6 and ACC are 0.8861,0.8396,0.8948 which are at least higher 0.2 than SSD YOLOv3,Faster R-CNN.It can be seen from the comparative experiment that this method for power towers detection has great application potential.

关 键 词：目标检测 遥感影像 可变形网络 迁移学习 Faster R-CNN 

分 类 号：P227[天文地球—大地测量学与测量工程]