Physics-informed radar weak target detection via signal structural information-guided deep binary classification

Weak target detection in low-SNR and cluttered radar environments remains challenging because conventional CFAR detectors are sensitive to heterogeneous backgrounds, while existing deep models often overlook radar-domain physical structures. To address this issue, this paper proposes a physics-informed weak-target detection framework based on signal structural information (SSI)-guided binary classification. A low-threshold CA-CFAR stage is first used to generate candidate target coordinates with high recall, after which an SSI extractor converts local range–Doppler neighborhoods into structure-preserving slices that retain the characteristic cross-shaped signatures induced by pulse compression and coherent Doppler integration. Based on these SSI slices, target detection is reformulated as a binary classification task. A dedicated classification network integrating anisotropic feature enhancement, dual-residual perception blocks, channel attention, and multiscale feature extraction is developed to capture the directional and scale-varying characteristics of target-related structures. Experiments on both simulated and real-measured datasets show that the proposed method achieves higher detection probability, stronger false-alarm suppression, and better robustness than conventional CFAR methods and representative deep-learning baselines. Ablation studies further verify the effectiveness of each component.