Synthetic Aperture Radar (SAR) image characteristics are highly susceptible to variations in the radar operation condition. Meanwhile, acquiring large amounts of SAR data under various imaging conditions is still a challenge in real application scenarios. Such sensitivity and scarcity bring an inadequately robust feature representation learning to recent data-hungry deep learning-based SAR Automatic Target Recognition (ATR) approaches. Considering the fact that physics-based electromagnetic simulated images could reproduce the image characteristics difference under various imaging conditions, we propose a simulation-aided domain adaptation technique to improve the generalization ability without extra measured SAR data. To be specific, We first build a surrogate feature alignment task using only simulated data based on a domain adaptation network. To mitigate the distribution shift problem between simulated and real data, we propose a category-level weighting mechanism based on SAR-SIFT similarity. This approach enhances surrogate feature alignment ability by re-weighting the simulated samples’ features in a category-level manner according to their similarities to the measured data. In addition, a meta-adaption optimization is designed to further reduce the sensitivity to the operation condition variation. We consider the recognition of the targets in simulated data across imaging conditions as an individual meta-task and adopt the multi-gradient descent algorithm to adapt the feature to different operation condition domains. We conduct experiments on two military vehicle datasets, MSTAR and SAMPLE-M with the aid of a simulated civilian vehicle dataset, SarSIM. The proposed method achieves state-of-the-art performance in extended operation conditions with 88.58% and 86.15% accuracy for variations in depression angle and resolution, outperforming our previous simulation-aided domain adaptation work TDDA. The code is available at https://github.com/ShShann/SA2FA-MAO.

中文翻译：

---

扩展作条件下 SAR ATR 的仿真辅助相似性感知特征对齐与元自适应优化


合成孔径雷达 （SAR） 图像特性极易受到雷达运行条件变化的影响。同时，在各种成像条件下获取大量 SAR 数据在实际应用场景中仍然是一个挑战。这种敏感性和稀缺性为最近基于数据的深度学习 SAR 自动目标识别 （ATR） 方法带来了不够稳健的特征表示学习。考虑到基于物理的电磁仿真图像在不同成像条件下可以再现图像特征差异，该文提出了一种仿真辅助域自适应技术，以便在没有额外测量 SAR 数据的情况下提高泛化能力。具体来说，我们首先仅使用基于域适应网络的模拟数据构建一个代理特征对齐任务。为了缓解模拟数据和真实数据之间的分布偏移问题，我们提出了一种基于 SAR-SIFT 相似性的类别级加权机制。这种方法通过根据模拟样本与测量数据的相似性，以类别级的方式对模拟样本的特征进行重新加权，从而提高了代理特征对齐能力。此外，还设计了元自适应优化，以进一步降低对作条件变化的敏感性。我们将跨成像条件下模拟数据中目标的识别视为一项单独的元任务，并采用多梯度下降算法使特征适应不同的工况域。我们借助模拟民用车辆数据集 SarSIM 对两个军用车辆数据集 MSTAR 和 SAMPLE-M 进行了实验。所提出的方法在扩展作条件下实现了 88.58% 和 86 的最先进的性能。俯角和分辨率变化的准确率为 15%，优于我们之前的仿真辅助域适应工作 TDDA。该代码可在 https://github.com/ShShann/SA2FA-MAO 获取。