基于深度学习的光学遥感图像目标检测方法综述

Deep Learning Based Object Detection in Optical RemoteSensing Images:A Survey

面向光学遥感图像的目标检测已经有较长时间的发展历史,近年来由于深度学习的发展而成为遥感领域的一个研究热点。目标检测的基础环节之一是图像特征提取,它对最终的检测效果有关键影响。早期目标检测器主要通过手动设计特征提取方法获得图像特征,往往较为复杂且难以提取图像的高层次特征。得益于深度学习技术的发展,目前目标检测器已经能够实现自动特征提取,取得了比传统方法更好的检测性能。对基于深度学习的光学遥感图像目标检测方法进行了总结和归纳:首先,回顾目标检测的发展历史;然后,介绍当前广泛使用的数据集;其次,将现有基于深度学习的光学遥感图像目标检测方法分为4类,并进行分析比较,包括基于人工提取候选区域的方法、基于候选区域生成网络的方法、基于回归的方法和其他方法;最后,分析现有光学遥感图像目标检测研究存在的问题,并对未来发展趋势进行了展望。

Object detection in optical remote sensing images has undergone years of development,and has become a hot research topic in recent years due to the development of deep learning.One fundamental step of object detection is image feature extraction,which has substantial impact on detection performance.Early object detectors extract image features using handcrafted feature extraction methods,which are sophisticated and difficult to extract high-level features.Current object detectors have benefited greatly from the success of deep learning and achieved great progress in automatic feature extraction,which makes them perform much better than traditional detectors.This paper presents a survey on deep learning-based object detection methods for optical remote sensing images.First,we brief the development history of object detection.Then,we introduce the publicly available datasets widely used in this filed.Following that,we classify the existing deep learning-based object detection methods for optical remote sensing images into four categories:methods based on handcrafted region proposal generation,methods based on region proposal network(RPN),methods based on regression,and the other methods.We also analyze and compare their merits and demerits.Finally,we highlight the open problems and future research directions.