基于卷积神经网络的无人机遥感影像农村建筑物目标检测  

Target Detection of Rural Buildings in UAV Remote Sensing Images Based on Convolutional Neural Network

作　　者：任媛媛 张显峰 马永建[1,2] 杨启原 汪传建 戴建国[1,2] 赵庆展 Ren Yuanyuan;Zhang Xianfeng;Ma Yongjian;Yang Qiyuan;Wang Chuanjian;Dai Jianguo;Zhao Qingzhan(College of Information Science and Technology,Shihezi University,Shihezi 832000,China;Xinjiang Production and Construction Corps Division of National Remote Sensing Center of China,Shihezi 832000,China;Institute of Remote Sensing and Geographic Information System,Peking University,Beijing 100871,China)

机构地区：[1]石河子大学信息科学与技术学院,新疆石河子832000 [2]国家遥感中心新疆兵团分部,新疆石河子832000 [3]北京大学遥感与地理信息系统研究所,北京100871

出　　处：《南京师范大学学报（工程技术版）》2019年第3期29-36,共8页Journal of Nanjing Normal University(Engineering and Technology Edition)

基　　金：国家重点研发计划(2017YFB0504203);国家自然科学基金(41461088);兵团科技计划(2016AB001、2017DB005)

摘　　要：将深度学习应用于遥感影像目标识别,提出基于卷积神经网络的无人机遥感影像农村建筑物的目标检测方法,用端到端的方式训练Faster R-CNN网络模型,并应用于农村建筑物的快速精确识别.该方法包括基于RPN网络的区域建议和基于Inception v2的卷积神经网络模型训练.为了训练和测试模型,通过无人机采集南疆地区的农村建筑物遥感影像,并人工标注建立了农村建筑物的数据集,在TensorFlow深度学习框架上通过对该数据集目标检测验证了模型.结果表明,基于改进的卷积神经网络目标检测方法对无人机遥感影像进行快速准确识别的总体精度超过90%,通过初始参数更新,模型收敛更快,对无人机遥感影像地物分类和目标识别具有一定的参考意义.With deep learning applied to object recognition of remote sensing images,a method of object detection for rural buildings based on convolution neural network is proposed in the paper.The improved Faster R-CNN network model is trained in an end-to-end way and applied to rural buildings with the rapid and accurate identification.Specifically,the method mainly includes region recommendation based on RPN network and convolutional neural network model training based on Inception v2.In order to train and test the improved model,the remote sensing images of rural buildings in southern Xinjiang Region are collected by UAV,and the data set of rural buildings is established by manual labeling.Finally,the model is validated by the object detection of the data set with TensorFlow deep learning framework.Experimental results show that the overall accuracy of fast and accurate recognition of UAV remote sensing images based on the improved convolution neural network object detection method exceeds 90%.By updating the initial parameters,the model converges faster,which has a certain reference value for the classification and object recognition of UAV remote sensing images.

关 键 词：建筑物 检测 无人机 深度学习 卷积神经网络 FASTER R-CNN 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]