摘要
深度学习技术的应用给SAR图像目标识别带来了大幅度的性能提升,但其对实际应用中车辆目标局部部件的变化适应能力仍有待加强。利用数据内在先验知识,在高维语义特征中学习其内在的低维子空间结构,可以提升分类模型在车辆目标变体条件下的泛化性能。本文基于目标特征的稀疏性,提出了一种稀疏先验引导卷积神经网络(Convolution Neural Network,CNN)学习的SAR目标识别方法(CNN-TDDL)。首先,该方法利用CNN提取SAR图像目标的高维语义特征。其次,通过稀疏先验引导模块,利用特征稀疏性,对目标特征内在的低维子空间结构进行学习。分类任务驱动的字典学习层(Task-Driven Dictionary Learning,TDDL)将目标特征的低维子空间以稀疏编码的形式表示,再利用非负弹性正则网增强了稀疏编码的稳定性,使稀疏编码不仅有效地表征目标的低维子空间结构,并且能够提取更具判别性的类别特征。基于运动和静止目标获取与识别(Moving and Stationary Target Acquisition and Recognition,MSTAR)数据集以及仿真和实测配对和标记实验(Synthetic and Measured Paired and Labeled Experiment,SAMPLE)数据集的实验表明,相比于传统字典学习方法和典型深度学习方法,CNN-TDDL在MSTAR标准操作条件(Standard Operating Conditions,SOC)下识别精度提升0.85%~5.28%,型号识别精度提升3.97%以上,表现出更好的泛化性能。特征可视化分析表明稀疏先验引导模块显著提升了异类目标特征表示的可分性。
The application of deep learning technology has greatly improved the performance of SAR image target recognition,but its adaptability to changes in local components of vehicle targets in practical applications still needs to be strengthened.Using the inherent prior knowledge of the data,learning its inherent low dimensional subspace structure from high-dimensional semantic features can improve the generalization performance of the classification model under the condition of vehicle target variants.Based on the sparsity of target features,this paper proposes a sparse priori guided convolution neural network(CNN)learning SAR target recognition method(CNN-TDDL).Firstly,CNN extracts high-dimensional semantic features of SAR image targets.Secondly,the sparse priori guidance module is used to learn the intrinsic low dimensional subspace structure of the target feature by using feature sparsity.The task driven dictionary learning(TDDL)layer of classification task represents the low dimensional subspace of the target feature in the form of sparse coding,and then uses the non negative elastic regular network to enhance the stability of sparse coding,so that sparse coding not only effectively represents the low dimensional subspace structure of the target,but also can extract more discriminative category features.Experiments based on moving and stationary target acquisition and recognition(MSTAR)datasets and synthetic and measured paired and labeled experiment(SAMPLE)datasets show that compared with traditional dictionary learning methods and typical deep learning methods,The identification accuracy of CNN-TDDL is improved by 0.85%~5.28%under MSTAR standard operating conditions(SOC),and the model identification accuracy is improved by more than 3.97%,showing better generalization performance.The feature visualization analysis shows that sparse prior guidance module significantly improves the separability of feature representation of heterogeneous targets.
作者
康志强
张思乾
封斯嘉
冷祥光
计科峰
KANG Zhiqiang;ZHANG Siqian;FENG Sijia;LENG Xiangguang;JI Kefeng(State Key Laboratory of Complex Electromagnetic Environment Effects on Electronics and Information System,College of Electronic Science,National University of Defense Technology,Changsha,Hunan 410073,China)
出处
《信号处理》
CSCD
北大核心
2023年第4期737-750,共14页
Journal of Signal Processing
基金
国家自然科学基金(62001480)
湖南省自然科学基金(2021JJ40684)。
关键词
合成孔径雷达
目标识别
卷积神经网络
稀疏先验
弹性正则网
synthetic aperture radar
target recognition
convolution neural network
sparse prior
elastic regular net
