测绘通报 ›› 2019, Vol. 0 ›› Issue (3): 27-31.doi: 10.13474/j.cnki.11-2246.2019.0072

• 学术研究 • 上一篇    下一篇

结合显著图和深度学习的遥感影像飞机目标识别

刘相云, 龚志辉, 金飞, 杨光, 范炜康   

  1. 信息工程大学, 河南 郑州 450001
  • 收稿日期:2018-07-28 修回日期:2018-11-07 出版日期:2019-03-25 发布日期:2019-04-02
  • 作者简介:刘相云(1994-),男,硕士生,主要研究方向为模式识别和机器学习。E-mail:liu_xy1994@163.com
  • 基金资助:
    飞机信息智能解译技术研究(校立[2018]107号)

Aircraft target recognition in remote sensing image combined saliency map with deep learning

LIU Xiangyun, GONG Zhihui, JIN Fei, YANG Guang, FAN Weikang   

  1. Information Engineering University, Zhengzhou 450001, China
  • Received:2018-07-28 Revised:2018-11-07 Online:2019-03-25 Published:2019-04-02

摘要: 为准确快速识别高分辨率遥感影像中的飞机目标,提出了一种结合显著图和深度置信网络(DBN)的飞机目标识别算法。本文首先使用HC (直方图对比度)算法提取遥感影像中的显著目标;然后通过定位连通区域确定候选目标的位置;随后提取候选目标的颜色矩、Hu不变矩、灰度共生矩阵、Tamura纹理特征和边缘方向直方图;最后将归一化后的多特征融合数据应用到深度置信网络进行目标识别。试验结果表明,本文算法的检测率为98.46%,虚警率为5.20%。算法从多种底层图像特征出发,有效克服了单一特征描述能力不足的问题,提高了飞机目标识别能力及抗干扰能力。

关键词: 飞机目标识别, 显著图, 多特征融合, 深度置信网络, 遥感影像

Abstract: In order to identify the aircraft target in high resolution remote sensing image accurately and quickly,a kind of aircraft target recognition algorithm combined the saliency map with the deep belief networks is proposed.First,the salient object in the image is extracted by use of histogram-based contrast method;Second,the candidate target is located by locating connected region;Then the color moment,Hu invariant moments,Tamura texture features and the edge direction histogram of candidate target is extracted.Last,the normalized feature is applied to deep belief networks to recognize the target.Experimental results show that the detection rate of the algorithm is 98.46% and the false alarm rate is 5.20%.Multi-feature provides more information than single-feature,and the ability of target recognition and anti-interference is improved.

Key words: aircraft targets recognition, saliency map, multi-feature fusion, deep belief networks, remote sensing image

中图分类号: