利用R—FCN的光学卫星影像典型目标检测  

Detection of Typical Target in Optical Satellite Image Based on R -FCN

作　　者：刘松林[1,2] 李新涛 巩丹超[1,2] 郭浩 Liu Songlin;Li Xintao;Gong Danchao;Guo Hao(Xi＇an Research Institute of Surveying and Mapping,Xi＇an 710054,China;State Key Laboratory of Geo-Information Engineering,Xi＇an 710054,China)

机构地区：[1]西安测绘研究所,陕西西安710054 [2]地理信息工程国家重点实验室,陕西西安710054

出　　处：《测绘科学与工程》2018年第3期52-57,共6页Geomatics Science and Engineering

摘　　要：传统的目标检测方法主要依靠人工设计特征，难以适用于海量遥感图像的多类目标检测任务。本文针对高分辨率光学卫星遥感影像中的舰船、飞机、储存罐和桥梁等目标，提出了一种利用R—FCN网络的多类目标检测方法。首先通过人工判读在影像中标注兴趣目标，构建样本库；然后使用样本库对R—FCN网络进行训练；最后，设计了重叠裁切和总体非极大值抑制策略，利用训练好的R—FCN模型完成目标检测。本文充分利用了深度卷积特征，避免了人工设计特征问题。通过在高分二号卫星数据集上进行的对比实验结果表明，本文方法能够快速准确地对多类目标同时进行检测，具有较好的准确性和鲁棒性。Traditional target detection of remote sensing image methods rely on artificial design features, which is difficult to apply to multiple target detection of massive remote sensing images. In this paper, a typical target detection method using the R - FCN network is proposed according to the targets of ship, plane, storage and bridge in high - resolution optical satellite remote sensing images. Firstly, the object of interest is marked in the image through manual interpretation and the sample library is constructed. Then the R - FCN network is trained using this sample library. Finally, the overlapped clipping and total non - maximum suppression strategies are designed and target detection is completed based on the trained R - FCN model. In this paper, the deep convolution features are fully utilized to avoid the problem of artificial design features. Experimental resuhs based on GF - 2 satellite data sets show that the typical target detection can be completed simultaneously using the proposed algorithm quickly and accurately, and it is more accurate and robust.

关 键 词：目标检测 光学影像 R—FCN 非极大值抑制 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]