北京航空航天大学学报 ›› 2019, Vol. 45 ›› Issue (1): 159-173.doi: 10.13700/j.bh.1001-5965.2018.0239

• 论文 • 上一篇    下一篇

遥感图像飞机目标高效搜检深度学习优化算法

郭琳, 秦世引   

  1. 北京航空航天大学 自动化科学与电气工程学院, 北京 100083
  • 收稿日期:2018-04-27 修回日期:2018-09-03 出版日期:2019-01-20 发布日期:2019-01-28
  • 通讯作者: 秦世引 E-mail:qsy@buaa.edu.cn
  • 作者简介:郭琳,男,博士研究生。主要研究方向:深度学习、图像语义分割、目标的检测与识别等;秦世引,男,博士,教授,博士生导师。主要研究方向:图像处理、模式识别、智能优化控制等。
  • 基金资助:
    国家自然科学基金(U1435220,61731001)

Deep learning and optimization algorithm for high efficient searching and detection of aircraft targets in remote sensing images

GUO Lin, QIN Shiyin   

  1. School of Automation Science and Electrical Engineering, Beihang University, Beijing 100083, China
  • Received:2018-04-27 Revised:2018-09-03 Online:2019-01-20 Published:2019-01-28
  • Supported by:
    National Natural Science Foundation of China (U1435220,61731001)

摘要: 为了实现大幅面遥感图像中飞机目标的高效检测与准确定位,通过深度神经网络(DNN)的级联组合,提出了一种新颖的搜寻与检测相集成的飞机目标高效检测算法。首先,运用高性能的端到端DNN网络,进行停机坪与跑道区域的像素级高效精准分割,从而大幅度缩小飞机目标的搜索范围,以降低虚警发生概率,完成飞机目标候选检测区域的快速搜寻。然后,针对分割所得停机坪与跑道区域,借助手工数据集对YOLO网络模型进行迁移式强化训练,一方面可以弥补训练集在样本类型与数据规模上的不足,另一方面借助YOLO网络的强时效性优势对飞机目标的位置进行回归求解,可以显著提高飞机目标的检测效率。停机坪与跑道区域分割DNN网络在分割精度与时效性上具有显著优势,而迁移式强化训练YOLO网络不仅具有很高的检测效率,在检测精度上也能保持良好的性能。通过一系列综合实验与对比分析,验证了提出的搜寻与检测相集成的DNN级联组合式飞机目标高效检测算法的性能优势。

关键词: 深度学习, 深度神经网络, 停机坪与跑道分割, 飞机目标检测, 大幅面遥感图像

Abstract: In order to achieve high-performance detection and accurate positioning of aircraft targets in large-scale remote sensing images, in this paper, a kind of efficient aircraft target detection algorithm based on synthetic integration of searching and detection is presented. First, through the end-to-end deep neural networks (DNN), the efficient and accurate pixel-level segmentation of apron and runway area is achieved so that the searching range of aircraft targets is greatly narrowed and the probability of false alarm is also reduced effectively and the goal of high speed searching of aircraft targets candidate detection areas is achieved accordingly. In view of the segmented areas of apron and runway, the strategy of transfer reinforcement learning is employed to pre-trained YOLO networks with supervised signals of positive datasets by manual labelling. In this way, pre-trained networks can make up the deficiency of capacity of manual data sets, and the advantage of real-time property of YOLO networks can also be utilized to deal with the classification and locations of aircraft targets so as to achieve a satisfied solution of regression problems and promote the efficiency of detection significantly. It is obvious that the apron and runway segmentation with DNN networks can play important role in getting performance superiority of high precision and efficiency. Meanwhile, YOLO networks based on transfer reinforcement learning not only possess the characteristics of high efficiency, but also maintain the precision of detection at a high level. A series of comprehensive experiments and comparative analyses verify the effectiveness and good performance of the proposed searching and detection algorithm of aircraft targets with DNN cascade combination and synthetic integration.

Key words: deep learning, deep neural networks, apron and runway segmentation, aircraft target detection, large-scale remote sensing image

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发