基于深度学习的小目标检测技术发展  

Development of Small Target Detection Technology Based on Deep Learning

作　　者：赵景波[1] 杜保帅 ZHAO Jingbo;DU Baoshuai(Qingdao University of Technology,Qingdao 266000 China)

机构地区：[1]青岛理工大学,山东青岛266000

出　　处：《电光与控制》2023年第2期63-70,74,共9页Electronics Optics & Control

基　　金：国家自然科学基金(51475251);山东省自然科学基金(ZR2013FM014)。

摘　　要：相较于大中等目标,像素点小于32×32的小目标由于存在特征信息少、像素低等问题,检测精度较差,导致水下机器人、无人机等设施对小目标物体的分类与定位精度低。为了提高相关设施对小目标的检测精度,分析了通用的深度学习目标检测算法现状,阐述了阻碍小目标检测技术发展的问题,对提升小目标检测精度的数据增强、特征融合、分辨率提升、上下文信息等多种类型算法进行了重点分析,最后归纳总结了各算法在MS COCO数据集上的AP与APs性能,并对小目标检测的未来发展方向进行了展望。Compared with that of the medium or large targets,the detection accuracy of small targets with pixels less than 32×32 is poor due to low pixels and lack of feature information,which results in low accuracy of underwater robots,UAVs and other facilities in classification and positioning of small objects.In order to improve the accuracy of the relative facilities in small target detection,analysis is made on the current situation of general deep-learning target detection algorithms at first.Then,the problems hindering the development of the small target detection technology are expounded,and the algorithms of data enhancement,feature fusion,resolution enhancement and context information that can improve the accuracy of small target detection are analyzed.Finally,the performances of AP and APs of all the algorithms in MS COCO data set are summarized,and the future development direction of small target detection is prospected.

关 键 词：小目标检测 深度学习 特征融合 数据增强 上下文信息 

分 类 号：V271.4[航空宇航科学与技术—飞行器设计]