测绘通报 ›› 2018, Vol. 0 ›› Issue (1): 77-82.doi: 10.13474/j.cnki.11-2246.2018.0014

• 行业观察 • 上一篇    下一篇

基于全卷积网络的高分辨遥感影像目标检测

徐逸之1,2, 姚晓婧1, 李祥1,2, 周楠3, 胡媛1,2   

  1. 1. 中国科学院遥感与数字地球研究所, 北京 100094;
    2. 中国科学院大学, 北京 100049;
    3. 苏州中科天启遥感科技有限公司, 江苏 苏州 215163
  • 收稿日期:2017-04-05 出版日期:2018-01-25 发布日期:2018-02-05
  • 通讯作者: 李祥。E-mail:lixiang01@radi.ac.cn E-mail:lixiang01@radi.ac.cn
  • 作者简介:徐逸之(1993-),男,硕士生,研究方向为地理信息系统研究与应用、数据挖掘。E-mail:sysu_xuyizhi@163.com
  • 基金资助:

    江苏省测绘地理信息科研项目(JSCHKY201720);国家自然科学基金(41701438);科技基金性工作专项重点项目(2014FY210800)

Object Detection in High Resolution Remote Sensing Images Based on Fully Convolution Networks

XU Yizhi1,2, YAO Xiaojing1, LI Xiang1,2, ZHOU Nan3, HU Yuan1,2   

  1. 1. Institute of Remote Sensing and Digital Earth, CAS, Beijing 100094, China;
    2. University of Chinese Academy of Sciences, Beijing 100049, China;
    3. Imagesky Remote Sensing Technology Co. Ltd., Suzhou 215163, China
  • Received:2017-04-05 Online:2018-01-25 Published:2018-02-05

摘要:

目标检测是遥感图像分析处理中的研究热点之一,具有十分重要的科研和应用价值。传统遥感影像目标检测方法多使用人工构造的浅层次特征,结合支持向量机、随机森林、Adaboost等分类器进行目标识别,难以充分挖掘和利用影像中的深层特征。近年来,深度学习,特别是卷积神经网络在图像认知方面取得了巨大成功。在目标检测领域,以Faster R-CNN算法为代表的方法取得了突破性进展,检测精度大幅提高,检测速度达到了近实时的性能。但是,Faster R-CNN算法由于使用了感兴趣区域(RoI)池化层,各个RoI计算不共享,因此检测速度依然有待提高。R-FCN基于全卷积网络结构,同时采用位置敏感池化来引入平移变化,抵消全卷积网络造成的平移不变形问题,检测精度和效率都有了很大的提高。本文阐述了R-FCN算法原理,并运用于高分辨遥感影像目标检测分析了不同参数和网络结构对R-FCN检测效果的影响,比较了利用Fast R-CNN、Faster R-CNN和R-FCN 3种算法进行飞机识别的性能。试验结果表明,利用R-FCN进行飞机识别定位可以达到99.3%的准确率和每张图180 ms的检测速度。

关键词: 高分辨率遥感, 深度学习, 全卷积网络, R-FCN, 飞机检测

Abstract:

Object detection is one of the research hotspots in remote sensing image analysis and processing,and has very important research and application value.Traditional remote sensing image target detection method uses the shallow artificial feature,and combines with classifiers,such as support vector machine,random forest and Adaboost,to realize object detection task.In recent years,deep learning,especially convolution neural network (CNN),has achieved great success in image recognition.In the field of target detection,the Faster R-CNN algorithm has made a breakthrough,the detection accuracy has been greatly improved,and the detection procedure has achieved nearly real-time performance.However,the Faster R-CNN algorithm uses RoI (region of interests) pooling layer,the RoI-wise calculation is not shared,so the detection speed is still unsatisfying.R-FCN,however,is based on the fully convolution network structure,and the position-sensitive pool is used to tackle the translation variance problem,which is concealed by the convolution neural network,thus lead to performance improvement.In this paper,the principle of R-FCN is described and applied to the detection of high-resolution remote sensing images.Different parameters and network structure are analyzed to find the best configuration.The performance of three popular algorithms,including Fast R-CNN,Faster R-CNN and R-FCN are compared.Experimental results show that R-FCN can achieve 99.3% precision and achieve at a test-time speed of 180 ms per image.

Key words: high resolution remote sensing, deep learning, fully convolution networks, R-FCN, aircraft detection

中图分类号: