Simulation-Aided SAR Target Classification via Dual-Branch Reconstruction and Subdomain Alignment

Convolutional neural networks (CNNs) are widely used in image classification, but such methods often require massive labeled data as learning resources. On the one hand, synthetic aperture radar (SAR) image interpretation is difficult, resulting in the lack of large-scale data sets with high-quality labels. On the other hand, CNNs are not explainable enough to provide reliable and trusted application services for SAR target recognition. To solve the above problems, physics-based electromagnetic simulated images are used to alleviate the shortage of real data with annotations, and explainability analysis methods are introduced to explain the basis of network decision-making. To address the domain gap between simulated and measured data, we propose a novel network integrating dual-branch image reconstruction and subdomain alignment (DBRSA). The network completes the reconstruction of simulated and measured images through the domain-shared encoder and domain-specific decoders, thereby helping the encoder to learn feature extraction methods independent of labels. In addition, the network aligns the feature vectors of similar targets obtained from different domains according to the real or pseudo labels of the samples, so as to further improve the classification accuracy. The experimental results and model decision analysis results demonstrate that the proposed network can improve the performance reliably by reducing the attention to the background noise and increasing the attention to the shadows and contours of the target, effectively reducing the dependence on the number of sample labels in practical application scenarios.