Unsupervised Cross-Domain Radar Target Recognition Using Multilevel Alignment

Abstract

Deep learning-based automatic target recognition (ATR) for synthetic aperture radar (SAR) has made significant advancements in recent years. However, many challenges persist, particularly in cross-domain applications from simulation training to measurement recognition. Although the electromagnetic simulation can provide abundant labeled training data, the domain shift between simulation and measurement results in poor generalization performance. Current methods often aim to reduce this discrepancy without a comprehensive analysis of domain shift. We adopt a novel perspective by splitting the SAR ATR into three parts: input, feature extraction, and output to analyze the domain shift. Guided by this analysis, we propose a multilevel alignment cross-domain recognition (MACR) network designed to progressively mitigate domain shift at the input, feature, and output levels, ultimately achieving full-process domain alignment between simulation and measurement. First, the gap is bridged through mutual conversion, generating simulated-like and measured-like samples to reduce the domain shift at the input level. Subsequently, adversarial learning is employed to diminish domain shift at the feature level. Finally, cross-domain knowledge distillation and pseudolabel filtering enforce consistency regularization based on category consistency priors between unlabeled measured and simulated-like samples, reducing domain shift at the output level. Experiments conducted on the synthetic and measured paired labeled experiment (SAMPLE) and SAMPLE+ datasets demonstrate the effectiveness of the proposed MACR, achieving state-of-the-art (SOTA) performance on both datasets.