基于深度学习的多尺度导弹发射井目标检测  

Multi-scale large well building object detection based on deep learning

作　　者：孟曦婷 计璐艳 赵永超 杨炜暾 MENG Xiting;JI Luyan;ZHAO Yongchao;YANG Weitun(Key Laboratory of Technology in Geo-spatial Information Processing and Application System of CAS,Aerospace Information Research Institutue,Chinese Academy of Sciences,Beijing 100094,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院空天信息创新研究院中国科学院空间信息处理与应用系统技术重点实验室,北京100094 [2]中国科学院大学,北京100049

出　　处：《中国科学院大学学报（中英文）》2021年第6期800-808,共9页Journal of University of Chinese Academy of Sciences

基　　金：国防科工局高分重大专项(30-Y20A15-9003-17/18,06-Y20A17-9001-17/18,30-Y20A28-9004-15/17);国家自然科学基金委国家重大科研仪器研制项目(41427805)资助。

摘　　要：导弹发射井是重要的遥感目标,发射井目标检测的研究对国防事业意义重大。在数据层面,由于发射井样本数量少,目前没有可用于其目标检测的有效数据集,构建有效的数据集对相关领域研究有极大价值。在算法层面,遥感图像分辨率的不同导致发射井目标呈现多尺度的特性,这是解决发射井目标检测问题的难点之一。基于以上分析,首先利用Google Earth构建首个发射井目标检测的数据集,然后针对发射井目标检测任务设计有效的检测模型。本文的模型充分融合了目标的多尺度特征和上下文的信息,并通过级联网络多阶段检测目标,有效检测出多尺度导弹发射井目标,检测效果优于目前主流的算法。Large well building is important remote sensing object,and the research on object detection of large well buildings is of great significance to national defense.At the data level,due to the small number of large well buildings samples,there is currently no valid data set available for the object detection.Building effective datasets is of great value for the research in related fields.At the algorithm level,the different resolutions of the remote sensing images result in multi-scale characteristics of the large well buildings,which is one of the difficulties in solving the object detection problem.Based on the above analysis,firstly,this article built the first large well buildings object detection dataset using Google Earth.Then an effective detection model was designed for large well building object detection task.The model in this paper fully integrates the object’s multi-scale features and contextual information,and detects the object through the multi-stage cascade network.The model can effectively detect large well buildings,and the detection effect is better than the results of the current mainstream algorithms.

关 键 词：导弹发射井 多尺度 特征融合 上下文信息 

分 类 号：TP751.2[自动化与计算机技术—检测技术与自动化装置]