J4 ›› 2015, Vol. 42 ›› Issue (3): 154-160.doi: 10.3969/j.issn.1001-2400.2015.03.026

• 研究论文 • 上一篇    下一篇

一种深度神经网络SAR遮挡目标识别方法

李帅;许悦雷;马时平;倪嘉成;史鹤欢   

  1. (空军工程大学 航空航天工程学院,陕西 西安  710038)
  • 收稿日期:2013-12-30 出版日期:2015-06-20 发布日期:2015-07-27
  • 通讯作者: 李帅
  • 作者简介:李帅(1988-),男,空军工程大学硕士研究生,E-mail: lishuailishuai@163.com.
  • 基金资助:

    国家自然科学基金资助项目(61379104,61372167); 航空科学基金资助项目(20115896022)

New method for SAR occluded targets recognition using DNN

LI Shuai;XU Yuelei;MA Shiping;NI Jiacheng;SHI Hehuan   

  1. (Institute of Aeronautics and Astronautics Engineering, Air Force Engineering Univ., Xi'an  710038, China)
  • Received:2013-12-30 Online:2015-06-20 Published:2015-07-27
  • Contact: LI Shuai

摘要:

提出了一种利用深度神经网络的合成孔径雷达图像部分遮挡目标的特征提取和目标识别新方法.该方法首先对合成孔径雷达图像进行预处理,然后提取预处理后合成孔径雷达目标的小波域低频子带图像作为训练数据,最后利用深层稀疏编码模型进一步提取合成孔径雷达遮挡目标的有效特征向量作为目标的特征以完成目标识别.采用MSTAR数据库中的3类目标进行目标遮挡模拟及识别实验.结果表明,新方法可以综合利用遮挡目标的局部和整体结构信息以提高目标的正确识别率,是一种有效的合成孔径雷达遮挡目标特征提取和目标识别方法.

关键词: 合成孔径雷达, 目标识别, 遮挡目标, 深度学习, 深层稀疏编码

Abstract:

For synthetic aperture radar(SAR) partial occluded images feature extraction and target recognition, a new method based on deep neural networks(DNN) is proposed. After preprocessing original images, we extract low-frequency sub-band images of SAR images in the wavelet domain as training data, and finally make a further extraction of the occluded targets' feature with the deep sparse autoencoder model as the input vectors. Three types of target in MSTAR database are used to simulate target occlusion and recognition experiment. Experimental results prove that the correct recognition rate could be improved by taking advantage of both local and global information about occluded targets, and that the new method is effective for feature extraction and object recognition of occluded SAR targets.

Key words: synthetic aperture radar, target recognition, occluded targets, deep learning, deep sparse autoencoders