摘要
独特的拍摄视角和多变的成像高度使得遥感影像中包含大量尺寸极其有限的目标,如何准确有效地检测这些小目标对于构建智能的遥感图像解译系统至关重要。本文聚焦于遥感场景,对基于深度学习的小目标检测进行全面调研。首先,根据小目标的内在特质梳理了遥感影像小目标检测的3个主要挑战,包括特征表示瓶颈、前背景混淆以及回归分支敏感。其次,通过深入调研相关文献,全面回顾了基于深度学习的遥感影像小目标检测算法。选取3种代表性的遥感影像小目标检测任务,即光学遥感图像小目标检测、SAR图像小目标检测和红外图像小目标检测,系统性总结了3个领域内的代表性方法,并根据每种算法使用的技术思路进行分类阐述。再次,总结了遥感影像小目标检测常用的公开数据集,包括光学遥感图像、SAR图像及红外图像3种数据类型,借助于3种领域的代表性数据集SODA-A(small object detection datasets)、AIR-SARShip和NUAA-SIRST(Nanjing University of Aeronautics and Astronautics,single-frame infrared small target),进一步对主流的遥感影像目标检测算法在面对小目标时的性能表现进行横向对比及深入评估。最后,对遥感影像小目标检测的应用现状进行总结,并展望了遥感场景下小目标检测的发展趋势。
Remote sensing images are often captured from multiview and multiple altitudes,thereby comprising a mass of objects with limited sizes which significantly challenge current detection methods that can achieve outstanding performance on natural images.Moreover,how to precisely detect these small objects plays a crucial role in developing an intelligent interpretation system for remote sensing images.Focusing on the remote sensing images,this paper conducts a comprehensive survey for deep learning-based small object detection(SOD)can be reviewed and analyzed literately,including 1)features-represent bottlenecks,2)background confusion,and 3)branching-regressed sensitivity.Specially,one of the major bottlenecks is for objective's representation.It refers that the down-sample operations in the prevailing feature extractors can suppress the signals of small objects unavoidably,and the following detection is impaired further in terms of the weak representations.The detection of size-limited instances is also interference of the confusion between the objects and backgrounds and the sensitivity of regression branch.For the former,the representations of small objects tend to be contaminated in related to feature extraction-contextual factors,which may erase the discriminative information that plays a significant role in head network.And the sensitivity of regression branch in small object detection is derived from the low tolerance for bounding box perturbation,in which a slight deviation of a predicted box will cause a drastic drop on the intersection-over-union(IoU),which is generally adopted to evaluate the accuracy of localization.Furthermore,we review and analyze the literature of small object detection for remote sensing images in the deep-learning era.In detail,by systematically reviewing corresponding methods of three small object detection tasks,i.e.,SOD for optical remote sensing images,SOD for synthetic aperture radar(SAR)images and SOD for infrared images,an understandable taxonomy of the reviewed algorithms for
作者
袁翔
程塨
李戈
戴威
尹文昕
冯瑛超
姚西文
黄钟泠
孙显
韩军伟
Yuan Xiang;Cheng Gong;Li Ge;Dai Wei;Yin Wenxin;Feng Yingchao;Yao Xiwen;Huang Zhongling;Sun Xian;Han Junwei(School of Automation,Northwestern Polytechnical University,Xi'an 710021,China;Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China)
出处
《中国图象图形学报》
CSCD
北大核心
2023年第6期1662-1684,共23页
Journal of Image and Graphics
基金
国家自然科学基金项目(62136007)
陕西省杰出青年科学基金项目(2021JC-16)。
