基于FCNN和ICAE的SAR图像目标识别方法

喻玲娟; 王亚东; 谢晓春; 林赟; 洪文

doi:10.12000/JR18066

基于FCNN和ICAE的SAR图像目标识别方法

doi: 10.12000/JR18066

喻玲娟^1,2, ,,
王亚东^1,,
谢晓春^3,,
林赟^2,,
洪文^2,

①.
江西理工大学信息工程学院赣州 341000
②.
中国科学院电子学研究所北京 100190
③.
赣南师范大学物理与电子信息学院赣州 341000

基金项目: 国家自然科学基金（61431018，61501210，61571421），江西省自然科学基金（20161BAB202054），江西省教育厅科技项目（GJJ150684，GJJ170825）

详细信息

作者简介:
喻玲娟(1982–)，女，籍贯江西，博士，江西理工大学副教授，硕士生导师，中国科学院电子学研究所博士后，研究方向为合成孔径雷达信号处理。E-mail: lingjuanyusmile@163.com

王亚东(1993–)，男，籍贯江苏，江西理工大学在读硕士研究生，研究方向为合成孔径雷达自动目标识别。E-mail: wangyadong183@163.com

谢晓春(1975–)，男，籍贯江西，博士，赣南师范大学副教授，硕士生导师，研究方向为合成孔径雷达信号处理。E-mail: xiexiaochun@gnnu.cn

林　赟(1983–)，女，籍贯浙江，博士，中国科学院电子学研究所副研究员，硕士生导师，研究方向为合成孔径雷达3维成像技术、多角度SAR成像基础理论与方法研究。E-mail: ylin@mail.ie.ac.cn

洪　文(1968–)，女，籍贯陕西，博士，中国科学院电子学研究所研究员，博士生导师，主要研究方向为合成孔径雷达成像与系统及其应用、极化/干涉合成孔径雷达数据处理及应用、3维微波成像新概念新体制新方法等。E-mail: whong@mail.ie.ac.cn

通讯作者:
喻玲娟　 lingjuanyusmile@163.com

计量
- 文章访问数: 2867
- HTML全文浏览量: 1014
- PDF下载量: 410
- 被引次数: 0
出版历程
- 收稿日期: 2018-08-31
- 修回日期: 2018-10-20
- 网络出版日期: 2018-10-28

SAR ATR Based on FCNN and ICAE

Yu Lingjuan^{1,2
, ,},
Wang Yadong^1
,,
Xie Xiaochun^3
,,
Lin Yun^2
,,
Hong Wen^2
,

①.
School of Information Engineering, Jiangxi University of Science and Technology, Ganzhou 341000, China
②.
Institute of Electronics, Chinese Academy of Sciences, Beijing 100190, China
③.
School of Physics and Electronic Information, Gannan Normal University, Ganzhou 341000, China

Funds: The National Natural Science Foundation of China (61431018, 61501210, 61571421), The Natural Science Foundation of Jiangxi Province (20161BAB202054), The Science and Technology Project of Jiangxi Provincial Education Department (GJJ150684, GJJ170825)

摘要

摘要: 近年来，基于卷积神经网络(Convolutional Neural Network, CNN)的合成孔径雷达(Synthetic Aperture Radar, SAR)图像目标识别得到深入研究。全卷积神经网络(Fully Convolutional Neural Network, FCNN)是CNN结构上的改进，它比CNN能获得更高的识别率，但在训练过程中仍需要大量的带标签训练样本。该文提出一种基于FCNN和改进的卷积自编码器(Improved Convolutional Auto-Encoder, ICAE)的SAR图像目标识别方法，即先用ICAE无监督训练方式获得的编码器网络参数初始化FCNN的部分参数，后用带标签训练样本对FCNN进行训练。基于MSTAR数据集的十类目标分类实验结果表明，在不扩充带标签训练样本的情况下，该方法不仅能获得98.14%的平均正确识别率，而且具有较强的抗噪声能力。
- 合成孔径雷达 /
- 自动目标识别 /
- 全卷积神经网络 /
- 卷积自编码器 /
- 改进的卷积自编码器
Abstract: In recent years, Synthetic Aperture Radar (SAR) image target recognition based on the Convolutional Neural Network (CNN) has attracted a significant amount of attention. Fully CNN (FCNN) is a structural improvement of the CNN, which features a higher recognition rate than CNN, but it still requires a large number of labeled data in the training process. This study aims to propose a method of SAR image target recognition based on FCNN and Improved Convolutional Auto-Encoder (ICAE), where several parameters of FCNN are initialized by the parameters of the ICAE encoder. These parameters are obtained in the unsupervised training mode. Then, the FCNN is trained by the labeled training samples. The experimental results on 10 kinds of target classification based on the MSTAR datasets show that this method cannot only achieve an average of 98.14% correct recognition rate but also feature a strong anti-noise capability when the labeled training samples are unexpanded.
- Synthetic Aperture Radar (SAR) /
- Automatic target recognition /
- Fully Convolutional Neural Network (FCNN) /
- Convolutional Auto-Encoder (CAE) /
- Improved Convolutional Auto-Encoder (ICAE)

HTML全文

图 1 CAE结构示意图

Figure 1. The structure of CAE

下载: 全尺寸图片幻灯片

图 3 FCNN的结构和前向传播

Figure 3. The structure and forward propagation of FCNN

下载: 全尺寸图片幻灯片

图 2 基于FCNN和ICAE的识别方法

Figure 2. The recognition method based on FCNN and ICAE

下载: 全尺寸图片幻灯片

图 4 ICAE的结构和前向传播

Figure 4. The structure and forward propagation of ICAE

下载: 全尺寸图片幻灯片

图 5 10类地面军事目标光学图像及其SAR图像

Figure 5. Ten types of ground military targets: Optical images versus SAR images

下载: 全尺寸图片幻灯片

图 6 4种方法的平均正确识别率随带标签的训练样本扩充倍数的变化

Figure 6. The average correct recognition rate of the four methods varies with the multiples of labeled training samples

下载: 全尺寸图片幻灯片

图 7 加不同比例噪声的SAR图像

Figure 7. SAR images with noise of different proportions

下载: 全尺寸图片幻灯片

图 8 4种方法的平均正确识别率随噪声所占比例的变化

Figure 8. The average correct recognition rate of the four methods varies with the noise proportion

下载: 全尺寸图片幻灯片

图 9 不同信噪比的SAR图像

Figure 9. SAR images with different SNR

下载: 全尺寸图片幻灯片

图 10 4种方法的平均正确识别率随信噪比的变化

Figure 10. The average correct recognition rate of the four methods varies with SNR

下载: 全尺寸图片幻灯片

表 1 训练样本和测试样本数量

Table 1. The number of training and testing images

型号	训练集		测试集
型号	角度	数量	角度	数量
BMP-2	17°	233	15°	195
BTR-70	17°	233	15°	196
BTR-60	17°	256	15°	195
BRDM-2	17°	298	15°	274
T-72	17°	232	15°	196
T-62	17°	299	15°	273
2S1	17°	299	15°	274
D7	17°	299	15°	274
ZIL-131	17°	299	15°	274
ZSU-234	17°	299	15°	274

下载: 导出CSV

表 2 FCNN, ICAE, CNN和CAE的网络结构

Table 2. The network structures of FCNN, ICAE, CNN and CAE

FCNN	ICAE	CNN	CAE
Conv.16@3×3 Conv.16@3×3/stride=2	Conv.16@3×3 Conv.16@3×3/stride=2	Conv.16@3×3 Maxpooling@2×2	Conv.16@3×3 Maxpooling@2×2
Conv.32@3×3 Conv.32@3×3/stride=2	Conv.32@3×3 Conv.32@3×3/stride=2	Conv.32@3×3 Maxpooling@2×2	Conv.32@3×3 Maxpooling@2×2
Conv.64@3×3 Conv.64@3×3/stride=2	Conv.64@3×3 Conv.64@3×3/stride=2	Conv.64@3×3 Maxpooling@2×2	Conv.64@3×3 Maxpooling@2×2
Conv.128@3×3 Conv.128@3×3/stride=2	Conv.128@3×3 Conv.128@3×3/stride=2	Conv.128@3×3 Maxpooling@2×2	Conv.128@3×3 Maxpooling@2×2
Conv.10@4×4 Softmax	Unpooling@2×2 Deconv.64@3×3	Conv.10@4×4 Softmax	Unpooling@2×2 Deconv.64@3×3
	Unpooling@2×2 Deconv.32@3×3		Unpooling@2×2 Deconv.32@3×3
	Unpooling@2×2 Deconv.16@3×3		Unpooling@2×2 Deconv.16@3×3
	Deconv.1@3×3		Deconv.1@3×3

下载: 导出CSV

表 3 基于FCNN和ICAE的识别结果

Table 3. The recognition results based on FCNN and ICAE

目标型号	识别结果										正确识别率 (%)
目标型号	2S1	BMP-2	BRDM-2	BTR-60	BTR-70	D7	T-62	T-72	ZIL-131	ZSU-234	正确识别率 (%)
2S1	265	0	1	0	2	0	0	6	0	0	96.72
BMP-2	1	191	0	0	1	1	0	1	0	0	98.45
BRDM-2	0	0	270	3	0	0	0	0	1	0	98.54
BTR-60	0	0	2	186	0	0	1	0	2	4	95.38
BTR-70	3	0	0	0	193	0	0	0	0	0	98.47
D7	0	0	2	0	0	272	0	0	0	0	99.27
T-62	0	0	0	0	0	0	270	0	0	3	98.90
T-72	0	1	0	0	1	0	0	194	0	0	98.98
ZIL-131	0	0	0	0	0	1	2	0	270	1	98.54
ZSU-234	0	0	0	0	0	3	1	1	0	269	98.18
平均正确识别率(%)	98.14

下载: 导出CSV

表 4 基于不同方法的实验结果对比

Table 4. The comparison of experimental results based on different methods

目标型号	识别率(%)
目标型号	FCNN+ICAE	CNN+CAE	FCNN	CNN
2S1	96.72	95.99	94.16	96.72
BMP-2	97.95	97.44	96.41	94.36
BRDM-2	98.54	95.99	96.35	95.26
BTR-60	95.38	91.79	95.38	96.41
BTR-70	98.47	99.49	98.98	93.88
D7	99.27	96.35	98.54	97.81
T-62	98.90	97.07	90.84	93.41
T-72	98.98	98.98	98.47	97.96
ZIL-131	98.54	98.91	97.81	96.35
ZSU-234	98.18	98.91	99.64	94.89
平均正确识别率(%)	98.14	97.11	96.58	95.71

下载: 导出CSV

参考文献(24)

[1]	Wagner S A. SAR ATR by a combination of convolutional neural network and support vector machines[J]. IEEE Transactions on Aerospace and Electronic Systems, 2016, 52(6): 2861–2872. DOI: 10.1109/TAES.2016.160061
[2]	Chen S Z, Wang H P, Xu F, et al. Target classification using the deep convolutional networks for SAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2016, 54(8): 4806–4817. DOI: 10.1109/TGRS.2016.2551720
[3]	Zhang Z M, Wang H P, Xu F, et al. Complex-valued convolutional neural network and its application in polarimetric SAR image classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2017, 55(12): 7177–7188. DOI: 10.1109/TGRS.2017.2743222
[4]	Furukawa H. Deep learning for target classification from SAR imagery: Data augmentation and translation invariance[J]. IEICE Technical Report, 2017, 117(182): 13–17.
[5]	徐丰, 王海鹏, 金亚秋. 深度学习在SAR目标识别与地物分类中的应用[J]. 雷达学报, 2017, 6(2): 136–148. DOI: 10.12000/JR16130 Xu Feng, Wang Haipeng, and Jin Yaqiu. Deep learning as applied in SAR target recognition and terrain classification[J]. Journal of Radars, 2017, 6(2): 136–148. DOI: 10.12000/JR16130
[6]	Morgan D A E. Deep convolutional neural networks for ATR from SAR imagery[C]. Proceedings of SPIE 9475, Algorithms for Synthetic Aperture Radar Imagery XXII, Baltimore, Maryland, United States, 2015: 94750F.
[7]	Profeta A, Rodriguez A, and Clouse H S. Convolutional neural networks for synthetic aperture radar classification[C]. Proceedings of SPIE 9843, Algorithms for Synthetic Aperture Radar Imagery XXIII, Baltimore, Maryland, United States, 2016: 98430M.
[8]	Wilmanski M, Kreucher C, and Lauer J. Modern approaches in deep learning for SAR ATR[C]. Proceedings of SPIE 9843, Algorithms for Synthetic Aperture Radar Imagery XXIII, Baltimore, Maryland, United States, 2016: 98430N.
[9]	Wagner S. Combination of convolutional feature extraction and support vector machines for radar ATR[C]. Proceedings of the 17th International Conference on Information Fusion, Salamanca, Spain, 2014: 1–6.
[10]	Pei J F, Huang Y L, Huo W B, et al. SAR automatic target recognition based on multiview deep learning framework[J]. IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(4): 2196–2210. DOI: 10.1109/TGRS.2017.2776357
[11]	Lin Z, Ji K F, Kang M, et al. Deep convolutional highway unit network for SAR target classification with limited labeled training data[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(7): 1091–1095. DOI: 10.1109/LGRS.2017.2698213
[12]	Chen S Z and Wang H P. SAR target recognition based on deep learning[C]. Proceedings of 2014 International Conference on Data Science and Advanced Analytics, Shanghai, China, 2015: 541–547.
[13]	El Housseini A, Toumi A, and Khenchaf A. Deep learning for target recognition from SAR images[C]. Proceedings of 2017 Seminar on Detection Systems Architectures and Technologies, Algiers, Algeria, 2017: 1–5.
[14]	田壮壮, 占荣辉, 胡杰民, 等. 基于卷积神经网络的SAR图像目标识别研究[J]. 雷达学报, 2016, 5(3): 320–325. DOI: 10.12000/JR16037 Tian Zhuangzhuang, Zhan Ronghui, Hu Jiemin, et al. SAR ATR based on convolutional neural network[J]. Journal of Radars, 2016, 5(3): 320–325. DOI: 10.12000/JR16037
[15]	Springenberg J T, Dosovitskiy A, Brox T, et al.. Striving for simplicity: The all convolutional net[OL]. arXiv preprint arXiv:1412.6806, 2015.
[16]	Hinton G E and Salakhutdinov R R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786): 504–507. DOI: 10.1126/science.1127647
[17]	康妙, 计科峰, 冷祥光, 等. 基于栈式自编码器特征融合的SAR图像车辆目标识别[J]. 雷达学报, 2017, 6(2): 167–176. DOI: 10.12000/JR16112 Kang Miao, Ji Kefeng, Leng Xiangguang, et al. SAR target recognition with feature fusion based on stacked autoencoder[J]. Journal of Radars, 2017, 6(2): 167–176. DOI: 10.12000/JR16112
[18]	赵飞翔, 刘永祥, 霍凯. 基于栈式降噪稀疏自动编码器的雷达目标识别方法[J]. 雷达学报, 2017, 6(2): 149–156. DOI: 10.12000/JR16151 Zhao Feixiang, Liu Yongxiang, and Huo Kai. Radar target recognition based on stacked denoising sparse autoencoder[J]. Journal of Radars, 2017, 6(2): 149–156. DOI: 10.12000/JR16151
[19]	Masci J, Meier U, Cireşan D, et al.. Stacked convolutional auto-encoders for hierarchical feature extraction[C]. Proceedings of the 21st International Conference on Artificial Neural Networks on Artificial Neural Networks and Machine Learning – ICANN 2011, Espoo, Finland, 2011: 52–59.
[20]	Vincent P, Larochelle H, Bengio Y, et al.. Extracting and composing robust features with denoising autoencoders[C]. Proceedings of the 25th International Conference on Machine Learning, Helsinki, Finland, 2008: 1096–1103.
[21]	Rumelhart D E, Hinton G E, and Williams R J. Learning representations by back-propagating errors[J]. Nature, 1986, 323(6088): 533–536. DOI: 10.1038/323533a0
[22]	Klambauer G, Unterthiner T, Mayr A, et al.. Self-normalizing neural networks[C]. Proceedings of the 31st Neural Information Processing Systems, Long Beach, CA, United States, 2017: 972–981.
[23]	Kingma D P and Ba J L. Adam: A method for stochastic optimization[OL]. arXiv preprint arXiv:1412.6980, 2015.
[24]	Dong G G, Wang N, and Kuang G Y. Sparse representation of monogenic signal: With application to target recognition in SAR images[J]. IEEE Signal Processing Letters, 2014, 21(8): 952–956. DOI: 10.1109/LSP.2014.2321565