基于特征空间结构对齐域自适应的机载SAR自动目标识别方法

Airborne SAR Automatic Target Recognition Method Using Domain Adaptation Based on Feature Space Structure Alignment

由于机载合成孔径雷达(SAR)系统的快速发展和深度学习的广泛应用,基于卷积神经网络的SAR图像自动目标识别技术已经取得了长足的发展。然而,由于真实测量获得的SAR数据较少,难以满足深度学习算法对大量训练样本的需求,目前已经有研究利用仿真SAR图像弥补真实SAR图像样本较少的缺陷。由于仿真SAR图像和真实SAR图像之间的差异,当前的主流研究方法是通过领域自适应(DA)的方法,将真实图像和仿真图像映射到相同的特征子空间中,从而提取域不变特征。然而,当前结合DA和仿真图像的SAR自动目标识别算法只关注了不同域之间样本内特征分布的相似度,却忽视了样本间的特征分布也包含了一定程度的语义信息。为了解决上述问题,本文提出了一种基于特征空间结构对齐的DA算法,充分挖掘仿真SAR图像和真实SAR图像之间共享的语义信息,从而显著地提升了深度学习模型在少样本情境下的识别性能。经过充分的试验和分析,试验结果证明了本文所提出方法不仅有较高的识别准确率,而且具备较强的泛用性和鲁棒性。

Due to the rapid development of airborne Synthetic Aperture Radar(SAR)system and deep learning theory,the airborne SAR automatic target recognition technology based on convolutional neural network has made great progress.However,due to the small amount of SAR data obtained by real measurements,it is difficult to meet the requirement of deep learning algorithms for a large number of training samples.At present,there have been studies using simulated SAR images to make up for the defect of the small number of real SAR image samples.Due to the differences between simulated SAR images and real SAR images,the current mainstream research method is to map real images and simulated images into the same feature subspace through Domain Adaptation(DA),so as to extract domain invariant features.However,the current SAR ATR algorithm combining DA and simulation images only pays attention to the similarity of feature distribution of the single sample pair in different domains,but ignores that the feature distribution between sample pairs also contains a certain degree of semantic information.In order to solve the above problems,this paper proposes a DA algorithm based on feature space structure alignment to fully mine the semantic information shared between simulated SAR images and measured SAR images,thus significantly improving the recognition performance of deep learning models in the context of small samples.After sufficient experiment and analysis,the experimental results prove that our proposed method not only has high recognition accuracy,but also has strong universality and robustness.