基于多域雷达回波数据融合的海面小目标分类网络模型

赵子健; 许述文; 水鹏朗

doi:10.11999/JEIT240818

基于多域雷达回波数据融合的海面小目标分类网络模型

doi: 10.11999/JEIT240818

西安电子科技大学雷达信号处理全国重点实验室西安 710071

基金项目: 国家自然科学基金(62371382)

详细信息

作者简介:
赵子健：男，博士生，研究方向为海面小目标识别、深度学习、机器学习、目标检测等

许述文：男，教授，研究方向为雷达目标检测与识别、机器学习、时频分析、SAR图像处理等

水鹏朗：男，教授，研究方向为海杂波建模、雷达目标检测与识别、图像处理等

通讯作者:
水鹏朗　plshui@xidian.edu.cn

中图分类号: TN957.52
计量
- 文章访问数: 223
- HTML全文浏览量: 88
- PDF下载量: 33
- 被引次数: 0
出版历程
- 收稿日期: 2024-09-24
- 修回日期: 2025-02-21
- 网络出版日期: 2025-03-06

A Network Model for Surface Small Targets Classification Based on Multidomain Radar Echo Data Fusion

National Key Laboratory of Radar Signal Processing, Xidian University, Xi’an 710071, China

Funds: The National Natural Science Foundation of China (62371382)

摘要

摘要: 海面小目标识别是海事雷达监视任务中一个重要且具有挑战性的问题。由于海面小目标类型多样、环境复杂多变，对其进行有效分类存在较大困难。在高分辨体制雷达下，海面小目标通常只占据一或几个距离单元，缺乏足够的空间散射结构信息，因此目标的雷达截面积(RCS)起伏和径向速度变化成为分类的主要依据。为此，该文提出一种基于多域雷达回波数据融合的分类网络模型，用于海面小目标的分类任务。由于不同域的数据具有其特殊的物理意义，因此该文构建了时域LeNet(T-LeNet)神经网络模块和时频特征提取神经网络模块，分别从雷达海面回波信号的幅度序列和时频分布(TFD)即时频图中提取特征。其中幅度序列主要反映了目标RCS的起伏特性，而时频图不仅反映RCS起伏特性，还能体现目标径向速度的变化。最后，利用IPIX, CSIR数据库和自测的无人机数据集构建了包括4种海面小目标的数据集：锚定漂浮小球、漂浮船只、低空无人机(UAV)和移动的快艇。实验结果表明所提方法具有良好的识别能力。
- 海面小目标 /
- 目标分类 /
- 多特征融合 /
- 神经网络
Abstract: Objective Small target recognition on the sea surface is a critical and challenging task in maritime radar surveillance. The variety of small targets and the complexity of the sea surface environment make their classification difficult. Due to the small size of these targets, typically occupying only one or a few range cells under high-resolution radar systems, there is insufficient spatial scattering structure information for classification. The primary information for classification comes from the target’s Radar Cross Section (RCS) fluctuation and radial velocity change. This study proposes a classification network model based on multidomain radar echo data fusion, providing a theoretical foundation for small target recognition in complex sea surface environments. Methods A small marine target classification network model is proposed, based on multidomain radar echo data fusion, incorporating both time domain and time-frequency domain. Given that data from different domains hold distinct physical significance, a Time-domain LeNet (T-LeNet) neural network module and a time-frequency feature extraction neural network module are designed to extract features from the amplitude sequence and the Time-Frequency Distribution (TFD), respectively. The amplitude sequence primarily reflects the fluctuation characteristics of the target’s RCS, while the TFD captures both the RCS fluctuations and variations in the target’s radial velocity. By extracting deep information from small sea surface targets, effective differential features are obtained, leading to improved classification results. The advantages of the multidomain data fusion approach are validated through ablation experiments, where the amplitude sequence is fused with the input TFD, or the TFD is fused with the input amplitude sequence. Additionally, the effect of network depth on recognition performance is explored by using ResNet architectures with varying depths for time-frequency feature extraction. Results and Discussions A dataset containing four types of small sea surface targets is constructed using measured data to evaluate the effectiveness of the proposed method. Six evaluation metrics are used to assess the model’s classification ability. The experimental results show that when only the TFD is input, the best recognition performance is achieved by the ResNet18 network. This is due to ResNet18’s ability to prevent gradient vanishing and explosion through residual connections, enabling a deeper network capable of more effectively extracting differential features between targets. When only the amplitude sequence is input, the recognition performance of the T-LeNet network improves significantly compared to the performance with only the TFD input. Fusing the amplitude sequence with the T-LeNet network, based solely on the input of the TFD, leads to a notable increase in recognition performance. Thus, incorporating information from other domains, such as time-domain information (amplitude sequence), and extracting abstract features from one-dimensional data with T-LeNet, while also capturing deeper target features from multidomain and multidimensional aspects, significantly enhances the network’s recognition capability. The best recognition performance occurs when both the amplitude sequence and TFD are input using the ResNet18 network, achieving an accuracy of 97.21%, which represents a 17% improvement over the TFD-only input with the Vgg16 network (Table 3). The confusion matrix reveals that Class I and Class II targets are more accurately classified when using only the amplitude sequence, with average accuracy improvements of 5% and 40%, respectively, compared to the TFD-only input. Class IV targets are better classified when using only the TFD, with an average accuracy improvement of 5% compared to the amplitude sequence input. There is no significant difference in the accuracy of Class III targets (Fig. 5). Comparing the classification results of different ResNet networks shows that increasing the depth of the ResNet network does not significantly enhance recognition performance (Table 4). Analyzing the loss and accuracy of the various experiments in both the training and validation sets reveals that combining the T-LeNet network improves performance further. Specifically, the accuracy of AlexNet, Vgg16, and ResNet18 in the validation set improves by approximately 7%, 5%, and 4%, respectively, while the loss in both the training and validation sets decreases (Fig. 6). Conclusions This paper proposes a small sea surface target classification method based on Convolutional Neural Networks (CNN) and data fusion. The method considers both the time domain and time-frequency domain, leveraging their distinct physical significance. It constructs the T-LeNet network module and the time-frequency feature extraction network module to extract deep information from small sea surface targets across multiple domains and dimensions. The abstract features jointly extracted from the time domain and time-frequency domain are fused for multidomain and multidimensional classification. The experimental results demonstrate that the proposed method exhibits strong recognition capability for small sea surface targets.
- Small sea surface target /
- Target classification /
- Multi-feature fusion /
- Neural network

HTML全文

图 1 4种海面小目标

下载: 全尺寸图片幻灯片

图 2 4类目标的幅度序列和时频图示例

下载: 全尺寸图片幻灯片

图 3 目标识别网络流程图

下载: 全尺寸图片幻灯片

图 4 网络结构

下载: 全尺寸图片幻灯片

图 5 混淆矩阵

下载: 全尺寸图片幻灯片

图 6 不同实验在训练集和验证集的损失与准确率

下载: 全尺寸图片幻灯片

表 1 4类目标与其对应的雷达参数

目标类型	数据来源	距离分辨率(m)	重频(kHz)	载频(GHz)	极化方式	工作模式	波束宽度(°)	海况(级)	训练样本数	测试样本数
锚定漂浮小球	IPIX93	30	1	9.39	HH/HV/VH/VV	驻留模式	0.9	2/3	21940	6560
漂浮船只	IPIX98	30	1	9.39	HH/HV/VH/VV	驻留模式	0.9	/	12768	3416
低空无人机	灵山岛	3	4	10.00	HH/VV	驻留模式	1.1	2	7569	1 893
移动的快艇	CSIR	15	2.5/5	6.90	VV	跟踪模式	1.8	3	2920	964

下载: 导出CSV

表 2 混淆矩阵

		预测类别
		目标1	目标2	目标3	目标4
真实类别	目标1	T₁P₁	F₁P₂	F₁P₃	F₁P₄
	目标2	F₂P₁	T₂P₂	F₂P₃	F₂P₄
	目标3	F₃P₁	F₃P₂	T₃P₃	F₃P₄
	目标4	F₄P₁	F₄P₂	F₄P₃	T₄P₄

下载: 导出CSV

表 3 不同实验在6个评价指标下的分类结果

	准确率	误差	精确度	召回率	F1-measure	Kappa
时频图+AlexNet	0.7773	0.2227	0.8139	0.7912	0.8024	0.6403
时频图+Vgg16^[5]	0.8022	0.1978	0.8461	0.8202	0.8330	0.6792
时频图+ResNet18	0.8145	0.1855	0.8487	0.8238	0.8361	0.7006
幅度序列+T-LeNet	0.9250	0.0750	0.9145	0.9130	0.9138	0.8823
幅度序列+时频图+T-LeNet+AlexNet	0.9426	0.0574	0.9440	0.9528	0.9484	0.9106
幅度序列+时频图+T-LeNet+Vgg16	0.9549	0.0451	0.9558	0.9578	0.9568	0.9296
幅度序列+时频图+T-LeNet+ResNet18	0.9721	0.0279	0.9708	0.9776	0.9742	0.9567

下载: 导出CSV

表 4 不同ResNet网络在6个评价指标下的分类结果

	准确率	误差	精确度	召回率	F1-measure	Kappa
时频图+ResNet18	0.8145	0.1855	0.8487	0.8238	0.8361	0.7006
时频图+ResNet34	0.8245	0.1755	0.8677	0.8379	0.8525	0.7165
时频图+ResNet50	0.8202	0.1798	0.8619	0.8359	0.8487	0.7100
幅度序列+时频图+T-LeNet+ResNet18	0.9721	0.0279	0.9708	0.9776	0.9742	0.9567
幅度序列+时频图+T-LeNet+ResNet34	0.9726	0.0274	0.9729	0.9775	0.9752	0.9574
幅度序列+时频图+T-LeNet+ResNet50	0.9736	0.0264	0.9707	0.9777	0.9742	0.9589

下载: 导出CSV

表 5 网络的参数量、训练时间、测试时间和单个样本测试时间

网络	参数量(M)	训练时间(min)	测试时间(s)	单个样本测试时间(ms)
T-LeNet	8.4669	7.25	23.87	1.86
AlexNet	61.1048	87.13	46.97	3.66
Vgg16	138.3651	216.15	55.05	4.29
ResNet18	11.1786	120.12	45.94	3.58
ResNet34	21.2867	195.19	72.65	5.66
ResNet50	23.5162	330.33	136.33	10.62
T-LeNet+AlexNet	71.7922	95.37	51.46	4.01
T-LeNet+Vgg16	149.0489	233.67	59.67	4.65
T-LeNet+ResNet18	21.7429	130.32	50.69	3.95
T-LeNet+ResNet34	31.8511	203.62	85.37	6.65
T-LeNet+ResNet50	34.4677	342.09	145.39	11.33

下载: 导出CSV

参考文献(22)

[1]	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.
[2]	关键. 雷达海上目标特性综述[J]. 雷达学报, 2020, 9(4): 674–683. doi: 10.12000/JR20114. GUAN Jian. Summary of marine radar target characteristics[J]. Journal of Radars, 2020, 9(4): 674–683. doi: 10.12000/JR20114.
[3]	NI Jun, ZHANG Fan, YIN Qiang, et al. Random neighbor pixel-block-based deep recurrent learning for polarimetric SAR image classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 59(9): 7557–7569. doi: 10.1109/TGRS.2020.3037209.
[4]	LEE S J, LEE M J, KIM K T, et al. Classification of ISAR images using variable cross-range resolutions[J]. IEEE Transactions on Aerospace and Electronic Systems, 2018, 54(5): 2291–2303. doi: 10.1109/TAES.2018.2814211.
[5]	XU Shuwen, RU Hongtao, LI Dongchen, et al. Marine radar small target classification based on block-whitened time-frequency spectrogram and pre-trained CNN[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5101311. doi: 10.1109/TGRS.2023.3240693.
[6]	GUO Zixun and SHUI Penglang. Anomaly based sea-surface small target detection using K-nearest neighbor classification[J]. IEEE Transactions on Aerospace and Electronic Systems, 2020, 56(6): 4947–4964. doi: 10.1109/TAES.2020.3011868.
[7]	KUO B C, HO H H, LI C H, et al. A kernel-based feature selection method for SVM with RBF kernel for hyperspectral image classification[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2014, 7(1): 317–326. doi: 10.1109/JSTARS.2013.2262926.
[8]	YIN Qiang, CHENG Jianda, ZHANG Fan, et al. Interpretable POLSAR image classification based on adaptive-dimension feature space decision tree[J]. IEEE Access, 2020, 8: 173826–173837. doi: 10.1109/ACCESS.2020.3023134.
[9]	JIA Qingwei, DENG Tingquan, WANG Yan, et al. Discriminative label correlation based robust structure learning for multi-label feature selection[J]. Pattern Recognition, 2024, 154: 110583. doi: 10.1016/j.patcog.2024.110583.
[10]	ZHONG Jingyu, SHANG Ronghua, ZHAO Feng, et al. Negative label and noise information guided disambiguation for partial multi-label learning[J]. IEEE Transactions on Multimedia, 2024, 26: 9920–9935. doi: 10.1109/TMM.2024.3402534.
[11]	ZHAO Jie, LING Yun, HUANG Faliang, et al. Incremental feature selection for dynamic incomplete data using sub-tolerance relations[J]. Pattern Recognition, 2024, 148: 110125. doi: 10.1016/j.patcog.2023.110125.
[12]	ZOU Yizhang, HU Xuegang, and LI Peipei. Gradient-based multi-label feature selection considering three-way variable interaction[J]. Pattern Recognition, 2024, 145: 109900. doi: 10.1016/j.patcog.2023.109900.
[13]	SUN Xu, GAO Junyu, and YUAN Yuan. Alignment and fusion using distinct sensor data for multimodal aerial scene classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5626811. doi: 10.1109/TGRS.2024.3406697.
[14]	WU Xin, HONG Danfeng, and CHANUSSOT J. Convolutional neural networks for multimodal remote sensing data classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5517010. doi: 10.1109/TGRS.2021.3124913.
[15]	DUAN Guoxing, WANG Yunhua, ZHANG Yanmin, et al. A network model for detecting marine floating weak targets based on multimodal data fusion of radar echoes[J]. Sensors, 2022, 22(23): 9163. doi: 10.3390/s22239163.
[16]	Cognitive Systems Laboratory. The IPIX radar database[EB/OL]. http://soma.ece.mcmaster.ca/ipix, 2001.
[17]	The Defense, Peace, Safety, and Security Unit of the Council for Scientific and Industrial Research. The Fynmeet radar database[EB/OL]. http://www.csir.co.ca/small_boat_detection, 2007.
[18]	RICHARD C. Time-frequency-based detection using discrete-time discrete-frequency Wigner distributions[J]. IEEE Transactions on Signal Processing, 2002, 50(9): 2170–2176. doi: 10.1109/TSP.2002.801927.
[19]	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: 770–778. doi: 10.1109/CVPR.2016.90.
[20]	KAYED M, ANTER A, and MOHAMED H. Classification of garments from fashion MNIST dataset using CNN LeNet-5 architecture[C]. 2020 International Conference on Innovative Trends in Communication and Computer Engineering (ITCE), Aswan, Egypt, 2020: 238–243. doi: 10.1109/ITCE48509.2020.9047776.
[21]	KRIZHEVSKY A, SUTSKEVER I, and HINTON G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84–90. doi: 10.1145/3065386.
[22]	SIMONYAN K and ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[C]. The 3rd International Conference on Learning Representations, San Diego, USA, 2015.