小样本SAR目标的双重一致性因果识别方法

王陈炜; 罗思懿; 黄钰林; 裴季方; 张寅; 杨建宇

doi:10.11999/JEIT240140

小样本SAR目标的双重一致性因果识别方法

doi: 10.11999/JEIT240140

电子科技大学信息与通信工程学院成都 611731

基金项目: 四川省自然科学基金 (2023NSFSC1970)

详细信息

作者简介:
王陈炜：男，博士，研究方向为雷达目标识别、机器学习、雷达探测成像等

罗思懿：女，硕士，研究方向为雷达目标识别、机器学习、雷达探测成像等

黄钰林：男，教授，研究方向为雷达探测成像、雷达目标检测与识别、人工智能与机器学习等

裴季方：男，副研究员，研究方向为雷达目标识别、机器学习、雷达探测成像等

张寅：男，研究员，研究方向为新体制雷达、雷达信号处理、雷达成像等

杨建宇：男，教授，研究方向为雷达探测与成像、信号检测与估计等

通讯作者:
黄钰林　yulinhuang@uestc.edu.cn

中图分类号: TN958
计量
- 文章访问数: 130
- HTML全文浏览量: 40
- PDF下载量: 31
- 被引次数: 0
出版历程
- 收稿日期: 2024-03-06
- 修回日期: 2024-08-28
- 网络出版日期: 2024-09-01
- 刊出日期: 2024-10-30

A Causal Interventional SAR ATR Method with Limited Data via Dual Consistency

School of Information and Communication Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China

Funds: The Natural Science Foundation of Sichuan Province (2023NSFSC1970)

摘要

摘要: 在小样本条件下提升方法的泛化性能，是合成孔径雷达自动目标识别(SAR ATR)的重要研究方向。针对该方向中的基础理论问题，该文建立了一个SAR ATR因果模型，证明了SAR图像中背景、相干斑等干扰在充足样本条件下可以被忽略；但在小样本条件下，这些因素将成为识别中的混杂因子，在提取的SAR图像特征中引入虚假相关性，影响SAR ATR性能。为了甄别和消除这些特征中的虚假效应，该文提出一个基于双重一致性的小样本SAR ATR方法，其中双重一致性包括类内一致性掩码和效应一致性损失。首先，基于鉴别特征应具有类内一致和类间差异的原则，利用类内一致性掩码，捕获目标的类内一致鉴别特征，甄别出目标特征中的混淆部分，准确估计出干扰引入的虚假效应。其次，基于不变风险最小化的思想，利用效应一致性损失，将经验风险最小化数据量需求转变为对效应相似度的度量需求，降低虚假效应消除对数据量的需求，消除特征中的虚假效应。因而，所提基于双重一致性的小样本SAR ATR方法可实现特征提取中的真实因果，实现准确的识别性能。两个基准数据集上的识别实验，验证了该方法的合理性和有效性，可提升小样本条件下SAR目标识别的性能。
- 合成孔径雷达 /
- 自动目标识别 /
- 小样本 /
- 因果推断
Abstract: Improving the generalization performance of methods under limited sample conditions is an important research direction in Synthetic Aperture Radar Automatic Target Recognition (SAR ATR). Addressing the fundamental problem in this field, a causal model is established in this paper for SAR ATR, demonstrating that interferences in SAR images, such as background and speckle, can be neglected under sufficient sample conditions. However, under limited sample conditions, these factors become confounding variables, introducing spurious correlations into the extracted SAR image features and affecting the generalization of SAR ATR. To accurately identify and eliminate these spurious effects in the features, this paper proposes a limited-sample SAR ATR method via dual consistency, which includes an intra-class feature consistency mask and effect-consistency loss. Firstly, based on the principle that discriminative features should have intra-class consistency and inter-class differences, the intra-class feature consistency mask is used to capture the consistent discriminative features of the target, subtracting the confounded part in the target features, and identifying the spurious effects introduced by interferences. Secondly, based on the idea of invariant risk minimization, the effect-consistency loss transforms the data requirement of empirical risk minimization into a need for labeling the similarity among effects of different samples, reducing the data demand for eliminating spurious effects and removing the spurious effects in the features. Thus, the limited-sample SAR ATR method proposed in this paper achieves true causal feature extraction and accurate recognition performance. Experiments on two benchmark datasets validate the effectiveness of this method which can achieve superior performance of SAR target recognition with limited sample.
- Synthetic Aperture Radar (SAR) /
- Automatic target recognition /
- Limited data /
- Causal inference

HTML全文

图 1 有限SAR样本在ATR中引入的因果效应变化

下载: 全尺寸图片幻灯片

图 2 本文方法的总体框架

下载: 全尺寸图片幻灯片

表 1 MSTAR数据集下的训练集与测试集

类别	训练集		测试集
类别	数量	俯仰角	数量	俯仰角
BMP2-9563	233	17°	195	15°
BRDM2-E71	298		274
BTR60-7532	256		195
BTR70-c71	233		196
D7-92	299		274
2S1-b01	299		274
T62-A51	299		273
T72-132	232		196
ZIL131-E12	299		274
ZSU234-d08	299		274

下载: 导出CSV

表 2 MSTAR数据集不同训练样本数量下10类目标的识别性能(%)

类别	每类目标训练样本数
类别	5	10	20	25	30	40	60	80	100
BMP2	73.33	83.59	92.31	94.87	95.90	97.95	98.97	97.44	98.46
BRDM2	77.74	94.89	93.80	96.72	97.08	98.18	99.27	99.64	98.91
BTR60	83.59	86.67	91.79	94.36	94.87	95.90	98.97	96.92	97.44
BTR70	70.92	90.82	92.35	95.92	96.94	98.47	97.96	98.98	98.98
D7	68.98	83.58	92.70	95.26	95.99	97.45	98.54	99.64	99.64
2S1	81.75	80.29	93.80	95.99	97.08	98.54	98.91	99.64	98.91
T62	74.36	90.11	92.67	95.60	97.07	98.53	97.44	99.63	99.63
T72	89.80	86.22	91.84	95.41	96.94	98.47	98.98	98.98	98.98
ZIL131	68.25	85.77	93.07	96.72	97.81	98.54	98.91	99.64	99.64
ZSU234	74.45	80.29	93.80	96.72	97.08	98.18	98.54	99.64	99.64
平均值	76.13	86.38	93.00	96.07	97.01	98.24	98.79	99.32	99.35

下载: 导出CSV

表 3 OpenSARShip数据集中训练与测试集

类别	成像条件	训练样本数	测试样本数	总样本数
Bulk Carrier	VH 和 VV, C 波段分辨率=5～20 m 入射角=20°～45° 仰角=±11° Rg20 m×az22 m	200	475	675
Container Ship		200	811	1011
Tanker		200	354	554

下载: 导出CSV

表 4 OpenSARShip数据集不同训练样本数量下3类舰船目标的识别性能(%)

类别	每类训练样本数
类别	10	20	30	40	50	60	70	80	100	200
Bulk Carrier	65.89	63.58	65.68	75.16	60.21	66.11	68.21	65.26	72.84	73.47
Container Ship	71.27	75.46	75.96	76.57	82.49	79.04	80.02	85.20	83.35	90.75
Tanker	79.10	82.77	81.92	74.01	84.46	86.16	84.46	81.92	78.81	82.77
平均值	71.51	73.75	74.41	75.66	76.58	76.92	77.66	78.86	79.41	84.12

下载: 导出CSV

表 5 消融实验：20个训练样本下不同消融配置的识别性能(%)

方法配置	特征	效应	BMP2	BRDM2	BTR60	BTR70	D7	2S1	T62	T72	ZIL131	ZSU234	平均值
V1	×	×	80.51	75.91	74.87	80.10	73.72	76.28	78.75	82.14	84.31	86.86	79.59
V2	√	×	88.21	89.05	81.54	88.27	85.77	96.35	81.68	84.18	87.23	94.16	88.16
V3	√	√	92.31	93.80	91.79	92.35	92.70	93.80	92.67	91.84	93.07	93.80	93.00

下载: 导出CSV

表 6 MSTAR数据集下SAR目标识别性能对比(%)

方法		每类样本数
方法		10	20	40	55	80	110	165	220	All data
传统方法	PCA+SVM ^[14]	–	76.43	87.95	–	92.48	–	–	–	94.32
	ADaboost ^[14]	–	75.68	86.45	–	91.45	–	–	–	93.51
	DGM^[14]	–	81.11	88.14	–	92.85	–	–	–	96.07
数据增强	GAN-CNN ^[14]	–	81.80	88.35	–	93.88	–	–	–	97.03
数据增强	MGAN-CNN^[14]	–	85.23	90.82	–	94.91	–	–	–	97.81
新颖模型	Deep CNN ^[15]	–	77.86	86.98	–	93.04	–	–	–	95.54
	Simple CNN ^[16]	–	75.88	–	–	–	–	–	–	–
	Dens-CapsNet ^[17]	80.26	92.95	96.50	–	–	–	–	–	99.75
	ASC-MACN ^[18]	62.85	79.46	–	–	–	–	–	–	99.42
	本文方法	86.38	93.00	98.24	–	99.32	–	–	–	–

下载: 导出CSV

参考文献(18)

[1]	CURLANDER J C and MCDONOUGH R N. Synthetic Aperture Radar: Systems and Signal Processing[M]. New York: Wiley, 1991.
[2]	KREUCHER C. SAR-ATR using EO-based deep networks[C]. 2023 IEEE Radar Conference (RadarConf23), San Antonio, USA, 2023: 1–5. doi: 10.1109/RadarConf2351548.2023.10149584.
[3]	MAO Deqing, YANG Jianyu, ZHANG Yongchao, et al. Angular superresolution of real aperture radar with high-dimensional data: Normalized projection array model and adaptive reconstruction[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5117216. doi: 10.1109/TGRS.2022.3203131.
[4]	MAO Deqing, YANG Jianyu, TUO Xingyu, et al. Angular superresolution of real aperture radar for target scale measurement using a generalized hybrid regularization approach[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5109314. doi: 10.1109/TGRS.2023.3315310.
[5]	CHOI J H, LEE M J, JEONG N H, et al. Fusion of target and shadow regions for improved SAR ATR[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5226217. doi: 10.1109/TGRS.2022.3165849.
[6]	HUANG Zhongling, YAO Xiwen, LIU Ying, et al. Physically explainable CNN for SAR image classification[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2022, 190: 25–37. doi: 10.1016/j.isprsjprs.2022.05.008.
[7]	WANG Chenwei, LUO Siyi, PEI Jifang, et al. Crucial feature capture and discrimination for limited training data SAR ATR[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2023, 204: 291–305. doi: 10.1016/j.isprsjprs.2023.09.014.
[8]	FENG Sijia, JI Kefeng, WANG Fulai, et al. PAN: Part attention network integrating electromagnetic characteristics for interpretable SAR vehicle target recognition[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5204617. doi: 10.1109/TGRS.2023.3256399.
[9]	WANG Chenwei, LUO Siyi, HUANG Yulin, et al. SAR ATR method with limited training data via an embedded feature augmenter and dynamic hierarchical-feature refiner[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5216215. doi: 10.1109/TGRS.2023.3314501.
[10]	PEARL J. Causal inference in statistics: An overview[J]. Statistics Surveys, 2009, 3: 96–146. doi: 10.1214/09-SS057.
[11]	SCHÖLKOPF B, JANZING D, PETERS J, et al. On causal and anticausal learning[C]. Proceedings of the 29th International Conference on Machine Learning, Edinburgh, UK, 2012.
[12]	GREENLAND S, PEARL J, and ROBINS J M. Causal diagrams for epidemiologic research[J]. Epidemiology, 1999, 10(1): 37–48. doi: 10.1097/00001648-199901000-00008.
[13]	ZHANG Tianwen, ZHANG Xiaoling, KE Xiao, et al. HOG-ShipCLSNet: A novel deep learning network with hog feature fusion for SAR ship classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5210322. doi: 10.1109/TGRS.2021.3082759.
[14]	ZHENG Ce, JIANG Xue, and LIU Xingzhao. Semi-supervised SAR ATR via multi-discriminator generative adversarial network[J]. IEEE Sensors Journal, 2019, 19(17): 7525–7533. doi: 10.1109/JSEN.2019.2915379.
[15]	MORGAN D A E. Deep convolutional neural networks for ATR from SAR imagery[C]. SPIE 9475, Algorithms for Synthetic Aperture Radar Imagery XXII, Baltimore, USA, 2015: 94750F. doi: 10.1117/12.2176558.
[16]	LI Yibing, LI Xiang, SUN Qian, et al. SAR image classification using CNN embeddings and metric learning[J]. IEEE Geoscience and Remote Sensing Letters, 2022, 19: 4002305. doi: 10.1109/LGRS.2020.3022435.
[17]	WANG Quan, XU Haixia, YUAN Liming, et al. Dense capsule network for SAR automatic target recognition with limited data[J]. Remote Sensing Letters, 2022, 13(6): 533–543. doi: 10.1080/2150704X.2022.2044089.
[18]	FENG Sijia, JI Kefeng, WANG Fulai, et al. Electromagnetic Scattering Feature (ESF) module embedded network based on ASC model for robust and interpretable SAR ATR[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5235415. doi: 10.1109/TGRS.2022.3208333.