一种基于EfficientNet与BiGRU的多角度SAR图像目标识别方法

赵鹏菲; 黄丽佳

doi:10.12000/JR20133

一种基于EfficientNet与BiGRU的多角度SAR图像目标识别方法

DOI: 10.12000/JR20133

赵鹏菲^{1, 2, 3,},
黄丽佳^{1, 2, ,}

1.
中国科学院空天信息创新研究院北京 100094
2.
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 100190
3.
中国科学院大学北京 100049

基金项目: 国家自然科学基金(61991420, 62022082)，中科院青促会专项支持

详细信息

作者简介:
赵鹏菲(1996–)，男，硕士生，研究方向为合成孔径雷达图像分析

黄丽佳(1984–)，女，博士，研究员，硕士生导师，研究方向为合成孔径雷达信号处理与图像分析

通讯作者:
黄丽佳 iecas8huanglijia@163.com

责任主编：林赟 Corresponding Editor: LIN Yun
中图分类号: TP753
计量
- 文章访问数: 3470
- HTML全文浏览量: 1274
- PDF下载量: 271
- 被引次数: 0
出版历程
- 收稿日期: 2020-10-26
- 修回日期: 2020-12-21
- 网络出版日期: 2021-01-07
- 刊出日期: 2021-12-28

Target Recognition Method for Multi-aspect Synthetic Aperture Radar Images Based on EfficientNet and BiGRU

ZHAO Pengfei^{1, 2, 3
,},
HUANG Lijia^{1, 2
, ,}

1.
Aerospace Information Research Institutue, Chinese Academy of Sciences, Beijing 100094, China
2.
Key Laboratory of Technology in Geo-spatial Information Processing and Application System, Chinese Academy of Sciences, Beijing 100190, China
3.
University of Chinese Academy of Sciences, Beijing 100049, China

Funds: The National Natural Science Foundation of China (61991420, 62022082), Special Support of Youth Innovation Promotion Association Chinese Academy of Sciences

More Information

Corresponding author: HUANG Lijia, iecas8huanglijia@163.com

摘要

摘要: 合成孔径雷达(SAR)的自动目标识别(ATR)技术目前已广泛应用于军事和民用领域。SAR图像对成像的方位角极其敏感，同一目标在不同方位角下的SAR图像存在一定差异，而多方位角的SAR图像序列蕴含着更加丰富的分类识别信息。因此，该文提出一种基于EfficientNet和BiGRU的多角度SAR目标识别模型，并使用孤岛损失来训练模型。该方法在MSTAR数据集10类目标识别任务中可以达到100%的识别准确率，对大俯仰角(擦地角)下成像、存在版本变体、存在配置变体的3种特殊情况下的SAR目标分别达到了99.68%, 99.95%, 99.91%的识别准确率。此外，该方法在小规模的数据集上也能达到令人满意的识别准确率。实验结果表明，该方法在MSTAR的大部分数据集上识别准确率均优于其他多角度SAR目标识别方法，且具有一定的鲁棒性。
- 合成孔径雷达 /
- 自动目标识别 /
- 多角度识别 /
- EfficientNet
Abstract: Automatic Target Recognition (ATR) in Synthetic Aperture Radar (SAR) has been extensively applied in military and civilian fields. However, SAR images are very sensitive to the azimuth of the images, as the same target can differ greatly from different aspects. This means that more reliable and robust multiaspect ATR recognition is required. In this paper, we propose a multiaspect ATR model based on EfficientNet and BiGRU. To train this model, we use island loss, which is more suitable for SAR ATR. Experimental results have revealed that our proposed method can achieve 100% accuracy for 10-class recognition on the Moving and Stationary Target Acquisition and Recognition (MSTAR) database. The SAR targets in three special imaging cases with large depression angles, version variants, and configuration variants reached recognition accuracies of 99.68%, 99.95%, and 99.91%, respectively. In addition, the proposed method achieves satisfactory accuracy even with smaller datasets. Our experimental results show that our proposed method outperforms other state-of-the-art ATR methods on most MSTAR datasets and exhibits a certain degree of robustness.
- Synthetic Aperture Radar (SAR) /
- Automatic Target Recognition (ATR) /
- Multi-aspect SAR /
- EfficientNet

HTML全文

图 1 多角度SAR目标识别网络结构图

Figure 1. Multi-aspect SAR ATR framework

下载: 全尺寸图片幻灯片

图 2 GRU结构示意图

Figure 2. The structure of GRU

下载: 全尺寸图片幻灯片

图 3 BiGRU结构示意图

Figure 3. The structure of BiGRU

下载: 全尺寸图片幻灯片

图 4 不同方位角、同一目标的SAR图像

Figure 4. SAR images of the same target with different azimuth

下载: 全尺寸图片幻灯片

图 5 多角度图像序列构造示意图

Figure 5. Schematic diagram of multi-angle image sequence structure

下载: 全尺寸图片幻灯片

表 1 EfficientNet-B0网络结构

Table 1. EfficientNet-B0 network structure

阶段	模块	输出尺寸	层数
1	Conv3×3	16×32×32	1
2	MBConv1, k3×3	24×32×32	1
3	MBConv6, k3×3	40×16×16	2
4	MBConv6, k5×5	80×8×8	2
5	MBConv6, k3×3	112×8×8	3
6	MBConv6, k5×5	192×4×4	3
7	MBConv6, k5×5	320×2×2	4
8	MBConv6, k3×3	1280×2×2	1
9	Conv1×1 & Pooling & FC	k	1

下载: 导出CSV

表 2 EfficientNet-B0与ResNet50网络对比

Table 2. Comparison of EfficientNet-B0 and ResNet50 networks

模型	参数量 (M)	FLOPS (B)	top1/top5准确率 (%)
EfficientNet-B0	5.3	0.39	77.3/93.5
ResNet50	26.0	4.10	76.0/93.0

下载: 导出CSV

表 3 图像序列L为4时，SOC数据集大小

Table 3. SOC dataset size when L=4

目标名称	训练集数量	测试集数量
2S1	1162	1034
BMP2	883	634
BRDM_2	1158	1040
BTR70	889	649
BTR60	978	667
D7	1162	1037
T62	1162	1032
T72	874	642
ZIL131	1162	1034
ZSU_234	1162	1040
合计	10592	8809

下载: 导出CSV

表 4 图像序列L为4时，EOC-1数据集大小

Table 4. EOC-1 dataset size when L=4

目标名称	训练集数量	测试集数量
2S1	1166	1088
BRDM_2	1162	1084
T72	913	1088
ZSU_234	1166	1088
合计	4407	4348

下载: 导出CSV

表 5 EOC-1, EOC-2与EOC-3数据集大小

Table 5. EOC-1, EOC-2 and EOC-3 dataset size

L	数据集	训练集总数	测试集总数
4	EOC-1	4407	4384
4	EOC-2	4473	9996
4	EOC-3	4473	12969
3	EOC-1	3307	3310
3	EOC-2	2889	7773
3	EOC-3	2889	10199
2	EOC-1	2202	2312
2	EOC-2	1934	5258
2	EOC-3	1934	6911

下载: 导出CSV

表 6 部分进行数据增广的数据集增广后大小

Table 6. The size of some data sets for data augmentation

L	数据集类型	训练集总数
4	EOC-1	17392
3	SOC	16032
3	EOC-1	13228
3	EOC-2&EOC-3	11544
2	SOC	16041
2	EOC-1	8808
2	EOC-2&EOC-3	7736

下载: 导出CSV

表 7 SOC实验中各参数设置

Table 7. Parameter in SOC experiment

名称	设置参数
Batch Size	32
优化器	Adam
Adam的学习率	0.001
Island Loss的优化器	SGD
SGD的学习率	0.5
Island Loss参数$ \lambda $	0.001
Island Loss参数$ { \lambda }_{1} $	10
Epochs	260

下载: 导出CSV

表 8 图像序列数L为4时，EOC-1混淆矩阵

Table 8. The EOC-1 confusion matrix when L=4

类型	S1	BRDM_2	T72	ZSU_234	Acc (%)
2S1	1076	2	10	0	98.90
BRDM_2	0	1084	0	0	100.00
T72	0	0	1088	0	100.00
ZSU_234	2	0	0	1086	99.82
平均值					99.68

下载: 导出CSV

表 9 图像序列数L为4时，各方法识别准确率在SOC与EOC-1数据集上对比

Table 9. Comparison of the recognition accuracy on SOC and EOC-1 dataset when L is 4

序号	方法	SOC			EOC-1
序号	方法	准确率 (%)	图像样本数量	图像序列样本数量	准确率 (%)	图像样本数量	图像序列样本数量
1	MVDCNN^[13]	98.52	690	43533	94.61	283	19705
2	MS-CNN^[15]	99.92	2747	2747	98.61	1128	1128
3	ResNet-LSTM^[16]	100.00	2000	7720	98.97	928	3614
4	本文方法	100.00	2747	10592	99.08	1128	4407
5	经过图像增广的本文方法	–	–	–	99.68	1128	17628

下载: 导出CSV

表 10 图像序列数L为3时，各方法准确率对比(%)

Table 10. Comparison of test accuracy when L=3 (%)

方法	SOC准确率	EOC-1准确率
MVDCNN^[13]	98.17	94.34
MS-CNN^[15]	99.88	97.48
本文方法	99.94	98.58

下载: 导出CSV

表 11 图像序列数L为2时，各方法准确率对比(%)

Table 11. Comparison of test accuracy when L=2 (%)

方法	SOC准确率	EOC-1准确率
MVDCNN^[13]	97.81	93.29
MS-CNN^[15]	99.84	96.69
本文方法	99.87	97.60

下载: 导出CSV

表 12 EOC-2数据集识别准确率对比(%)

Table 12. Comparison of accuracy on EOC-2 (%)

方法	L=4	L=3	L=2
MVDCNN^[13]	95.46	95.08	93.75
MS-CNN^[15]	100.00	100	99.67
本文方法	99.95	99.82	99.39

下载: 导出CSV

表 13 EOC-3数据集识别准确率对比(%)

Table 13. Comparison of accuracy on EOC-3 (%)

方法	L=4	L=3	L=2
MVDCNN^[13]	95.45	95.25	94.98
MS-CNN^[15]	99.58	99.08	98.71
本文方法	99.91	99.57	99.13

下载: 导出CSV

表 14 在缩减数据集上的识别准确率(%)

Table 14. Recognition accuracy on the reduced dataset (%)

数据集规模	5%	15%	50%
本文方法	95.98	99.72	99.93
ResNet-LSTM^[16]	93.97	99.37	99.58

下载: 导出CSV

表 15 消融实验结果

Table 15. Results of ablation experiments

序号	Center Loss	Island Loss	EfficientNet	BiGRU	准确率 (%)	提升 (%)
1					94.08	–
2	√				95.81	1.73
3		√			97.03	1.22
4		√	√		98.46	1.43
5		√	√	√	99.08	0.62

下载: 导出CSV

参考文献(22)

[1]	盖旭刚, 陈晋汶, 韩俊, 等. 合成孔径雷达的现状与发展趋势[J]. 飞航导弹, 2011(3): 82–86, 95. GAI Xugang, CHEN Jinwen, HAN Jun, et al. Development status and trend of synthetic aperture radar[J]. Aerodynamic Missile Journal, 2011(3): 82–86, 95.
[2]	张红, 王超, 张波, 等. 高分辨率SAR图像目标识别[M]. 北京: 科学出版社, 2009. ZHANG Hong, WANG Chao, ZHANG Bo, et al. Target Recognition in High Resolution SAR Images[M]. Beijing: Science Press, 2009.
[3]	MOREIRA A, PRATS-IRAOLA P, YOUNIS M, et al. A tutorial on synthetic aperture radar[J]. IEEE Geoscience and Remote Sensing Magazine, 2013, 1(1): 6–43. doi: 10.1109/MGRS.2013.2248301
[4]	王瑞霞, 林伟, 毛军. 基于小波变换和PCA的SAR图像相干斑抑制[J]. 计算机工程, 2008, 34(20): 235–237. doi: 10.3969/j.issn.1000-3428.2008.20.086 WANG Ruixia, LIN Wei, and MAO Jun. Speckle suppression for SAR image based on wavelet transform and PCA[J]. Computer Engineering, 2008, 34(20): 235–237. doi: 10.3969/j.issn.1000-3428.2008.20.086
[5]	CHEN Sizhe and WANG Haipeng. SAR target recognition based on deep learning[C]. 2014 International Conference on Data Science and Advanced Analytics, Shanghai, China, 2015.
[6]	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
[7]	CHEN Sizhe, WANG Haipeng, XU Feng, 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
[8]	FURUKAWA H. Deep learning for target classification from SAR imagery: Data augmentation and translation invariance[R]. SANE2017-30, 2017.
[9]	袁媛, 袁昊, 雷玲, 等. 一种同步轨道星机双基SAR成像方法[J]. 雷达科学与技术, 2007, 5(2): 128–132. doi: 10.3969/j.issn.1672-2337.2007.02.011 YUAN Yuan, YUAN Hao, LEI Ling, et al. An imaging method of GEO Spaceborne-Airborne Bistatic SAR[J]. Radar Science and Technology, 2007, 5(2): 128–132. doi: 10.3969/j.issn.1672-2337.2007.02.011
[10]	史洪印, 周荫清, 陈杰. 同步轨道星机双基地三通道SAR地面运动目标指示算法[J]. 电子与信息学报, 2009, 31(8): 1881–1885. SHI Hongyin, ZHOU Yinqing, and CHEN Jie. An algorithm of GEO spaceborne-airborne bistatic three-channel SAR ground moving target indication[J]. Journal of Electronics &Information Technology, 2009, 31(8): 1881–1885.
[11]	LI Zhuo, LI Chunsheng, YU Ze, et al. Back projection algorithm for high resolution GEO-SAR image formation[C]. 2011 IEEE International Geoscience and Remote Sensing Symposium, Vancouver, Canada, 2011: 336–339.
[12]	ZHANG Fan, HU Chen, YIN Qiang, et al. Multi-aspect-aware bidirectional LSTM networks for synthetic aperture radar target recognition[J]. IEEE Access, 2017, 5: 26880–26891. doi: 10.1109/ACCESS.2017.2773363
[13]	PEI Jifang, HUANG Yulin, HUO Weibo, 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
[14]	邹浩, 林赟, 洪文. 采用深度学习的多方位角SAR图像目标识别研究[J]. 信号处理, 2018, 34(5): 513–522. doi: 10.16798/j.issn.1003-0530.2018.05.002 ZOU Hao, LIN Yun, and HONG Wen. Research on multi-aspect SAR images target recognition using deep learning[J]. Journal of Signal Processing, 2018, 34(5): 513–522. doi: 10.16798/j.issn.1003-0530.2018.05.002
[15]	ZHAO Pengfei, LIU Kai, ZOU Hao, et al. Multi-stream convolutional neural network for SAR automatic target recognition[J]. Remote Sensing, 2018, 10(9): 1473. doi: 10.3390/rs10091473
[16]	ZHANG Fan, FU Zhenzhen, ZHOU Yongsheng, et al. Multi-aspect SAR target recognition based on space-fixed and space-varying scattering feature joint learning[J]. Remote Sensing Letters, 2019, 10(10): 998–1007. doi: 10.1080/2150704X.2019.1635287
[17]	TAN Mingxing and LE Q V. EfficientNet: Rethinking model scaling for convolutional neural networks[J]. arXiv: 1905.11946, 2019.
[18]	CHO K, VAN MERRIENBOER B, GULCEHRE C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[J]. arXiv: 1406.1078, 2014.
[19]	CAI Jie, MENG Zibo, KHAN A S, et al. Island loss for learning discriminative features in facial expression recognition[C]. The 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), Xi’an, China, 2018: 302–309.
[20]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016.
[21]	HOCHREITER S and SCHMIDHUBER J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735–1780. doi: 10.1162/neco.1997.9.8.1735
[22]	WEN Yandong, ZHANG Kaipeng, LI Zhifeng, et al. A discriminative feature learning approach for deep face recognition[C]. The 14th European Conference on Computer Vision – ECCV 2016, Amsterdam, The Netherlands, 2016.