基于GCN和CNN联合的SAR图像自动目标识别  

Automatic Target Recognition for SAR Images Based on the Combination of GCN and CNN

作　　者：秦基凯 刘峥[1] 谢荣[1] 冉磊[1] QIN Jikai;LIU Zheng;XIE Rong;RAN Lei(National Key Laboratory of Radar Signal Processing,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学雷达信号处理全国重点实验室,陕西西安710071

出　　处：《雷达科学与技术》2024年第6期587-595,共9页Radar Science and Technology

基　　金：国家自然科学基金(No.62001346)。

摘　　要：基于卷积神经网络(Convolutional Neural Network, CNN)的合成孔径雷达(Synthetic Aperture Radar,SAR)自动目标识别(Automatic Target Recognition, ATR)技术近些年来备受关注,已成为SAR图像解译领域的研究热点。然而,这类方法主要利用的是SAR图像的幅值信息,仅从局部区域中提取特征。鉴于SAR图像中的目标通常被视为散射中心的相干叠加,这些目标展现出复杂的结构和丰富的上下文信息。仅依靠CNN难以充分捕捉目标周围的全局信息,这可能会影响识别精度。因此,为了进一步提高识别性能,本研究引入图卷积网络(Graph Convolutional Network, GCN),提出一种结合GCN和CNN的SAR ATR方法。该方法首先利用传统CNN提取与SAR图像幅值相关的局部特征,接着通过构造图数据并应用GCN提取全局特征。此外,本研究还设计了多尺度GCN,通过融合不同尺度的特征来增强模型对图数据的学习能力。在模型训练阶段,采用标签平滑技术以缓解过拟合问题。通过端到端的训练策略,实现了GCN和CNN参数的联合优化,从而实现高精度的SAR图像目标识别。最终,通过在MSTAR和OpenSARship数据集上的实验表明,所提方法在识别性能上优于现有技术,并展现出卓越的泛化能力。The automatic target recognition(ATR)technology based on convolutional neural network(CNN)forsynthetic aperture radar(SAR)has attracted much attention in recent years, becoming a research hotspot in the field ofSAR image interpretation. However, these methods primarily utilize the amplitude information of SAR images and onlyextract features from local regions. Given that targets in SAR images are typically regarded as the coherent superpositionof scattering centers, these targets exhibit complex structures and rich contextual information. It is difficult to fully capturethe global information around the target by relying only on CNN, which may affect the recognition accuracy. Therefore,to further improve the recognition performance, this study introduces the graph convolutional network(GCN)andproposes a SAR ATR method combining GCN and CNN. This method first utilizes traditional CNN to extract local featuresrelated to the amplitude of SAR images, and then employs GCN to extract global features by constructing graphdata. Additionally, a multi‑scale GCN is designed to enhance the interpretation ability of the graph data by fusing featuresfrom different scales. During the model training phase, the label smoothing technique is employed to alleviate theoverfitting. Through an end‑to‑end training strategy, the joint optimization of GCN and CNN parameters ensures a highprecisionSAR image target recognition. Finally, experimental results on the MSTAR and OpenSARship datasets demonstratethat the proposed method outperforms the existing techniques in terms of recognition performance and exhibitssuperior generalization capability.

关 键 词：合成孔径雷达 图卷积网络 卷积神经网络 自动目标识别 多尺度GCN 

分 类 号：TN958[电子电信—信号与信息处理] TP183[电子电信—信息与通信工程]