基于卷积神经网络的光学遥感目标检测研究进展  

Research Progress on Optical Remote Sensing Object Detection Based on CNN

作　　者：张春晓[1] 鲍云飞[1] 马中祺 陶睿 李维[1] 田宇 ZHANG Chunxiao;BAO Yunfei;MA Zhongqi;TAO Rui;LI Wei;TIAN Yu(Beijing Institute of Space Mechanics&Electricity,Beijing 100094,China;State Grid Jibei Information&Telecommunication Company,Beijing 100053,China)

机构地区：[1]北京空间机电研究所,北京100094 [2]国网冀北电力有限公司信息通信分公司,北京100053

出　　处：《航天返回与遥感》2020年第6期45-55,共11页Spacecraft Recovery & Remote Sensing

摘　　要：光学遥感目标检测是一类特殊的目标检测领域,是光学遥感图像分析与应用的重要环节。文章围绕遥感目标检测的特点和难点,在自然图像目标检测网络基础上,重点分析了检测网络针对遥感领域的三方面适应性改进:数据量大、背景复杂度高、目标形态多样的适应性改进,对目标方向的适应性改进和遥感数据域迁移的适应性改进。此外,论文还给出了当前公开的光学遥感目标检测数据集及相关应用,最后总结了深度学习在光学遥感目标检测中的局限性和未来发展趋势,有助于光学目标检测技术的发展。Optical remote sensing object detection is a special type of object detection,which is an important content of optical remote sensing image analysis and application.Aiming at the characteristics and challenges of remote sensing object detection,and on the basis of the outstanding natural-image object detection networks,this paper focuses on the three aspects of adaptability improvement of the detection networks,involving the adaptability for large image size,high background complexity,and small targets,the adaptability for object rotations,and the adaptability for the domain migration.In addition,the paper presents the public optical remote sensing object detection data sets and related applications,and finally summarizes the limitations and future research directions of deep learning in optical remote sensing object detection.

关 键 词：目标检测 卷积神经网络 光学遥感 

分 类 号：P407.8[天文地球—大气科学及气象学]