基于图网络与不变性特征感知的SAR图像目标识别方法  

Target Recognition Method Based on Graph Structure Perception of Invariant Features for SAR Images

作　　者：曹婧宜 张扬[1] 尤亚楠 王亚敏[2] 杨峰 任维佳 刘军[1] CAO Jingyi;ZHANG Yang;YOU Ya’nan;WANG Yamin;YANG Feng;REN Weijia;LIU Jun(School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876,China;School of Electronic and Information Engineering,Beihang University,Beijing 100191,China;Changsha Tianyi Space Science and Technology Research Institute Co.,Ltd.,Changsha 410221,China)

机构地区：[1]北京邮电大学人工智能学院,北京100876 [2]北京航空航天大学电子信息工程学院,北京100191 [3]长沙天仪空间科技研究院有限公司,长沙410221

出　　处：《雷达学报(中英文)》2025年第2期366-388,共23页Journal of Radars

基　　金：国家重点研发计划(2023YFC3305901);国家自然科学基金(62101060)。

摘　　要：基于深度学习的合成孔径雷达(SAR)图像目标识别技术日趋成熟。然而,受散射特性、噪声干扰等影响,同类目标的SAR成像结果存在差异。面向高精度目标识别需求,该文将目标实体、生存环境及其交互空间中不变性特征的组合抽象为目标本质特征,提出基于图网络与不变性特征感知的SAR图像目标识别方法。该方法用双分支网络处理多视角SAR图像,通过旋转可学习单元对齐双支特征并强化旋转免疫的不变性特征。为实现多粒度本质特征提取,设计目标本体特征强化单元、环境特征采样单元、上下文自适应融合更新单元,并基于图神经网络分析其融合结果,构建本质特征拓扑,输出目标类别向量。该文使用t-SNE方法定性评估算法的类别辨识能力,基于准确率等指标定量分析关键单元及整体网络,采用类激活图可视化方法验证各阶段、各分支网络的不变性特征提取能力。该文所提方法在MSTAR车辆、SAR-ACD飞机、OpenSARShip船只数据集上的平均识别准确率分别达到了98.56%,94.11%,86.20%。实验结果表明,该算法具备在SAR图像目标识别任务中目标本质特征提取能力,在多类别目标识别方面展现出较高的稳健性。Synthetic Aperture Radar(SAR)image target recognition technology based on deep learning has matured.However,challenges remain due to scattering phenomenon and noise interference that cause significant intraclass variability in imaging results.Invariant features,which represent the essential attributes of a specific target class with consistent expressions,are crucial for high-precision recognition.We define these invariant features from the entity,its surrounding environment,and their combined context as the target’s essential features.Guided by multilevel essential feature modeling theory,we propose a SAR image target recognition method based on graph networks and invariant feature perception.This method employs a dualbranch network to process multiview SAR images simultaneously using a rotation-learnable unit to adaptively align dual-branch features and reinforce invariant features with rotational immunity by minimizing intraclass feature differences.Specifically,to support essential feature extraction in each branch,we design a featureguided graph feature perception module based on multilevel essential feature modeling.This module uses salient points for target feature analysis and comprises a target ontology feature enhancement unit,an environment feature sampling unit,and a context-based adaptive fusion update unit.Outputs are analyzed with a graph neural network and constructed into a topological representation of essential features,resulting in a target category vector.The t-Distributed Stochastic Neighbor Embedding(t-SNE)method is used to qualitatively evaluate the algorithm’s classification ability,while metrics like accuracy,recall,and F1 score are used to quantitatively analyze key units and overall network performance.Additionally,class activation map visualization methods are employed to validate the extraction and analysis of invariant features at different stages and branches.The proposed method achieves recognition accuracies of 98.56% on the MSTAR dataset,94.11%on SAR-ACD dataset,and 86.20% o

关 键 词：合成孔径雷达(SAR) 目标识别 不变性特征提取 本质特征 深度学习 

分 类 号：TN957.52[电子电信—信号与信息处理]