Physics-Based Synthetic Training for Automatic Target Detection in Radar Imaging

Representative labeled radar data are difficult to acquire across the target and scene variability needed for automatic detection. Physics-based simulation offers an alternative, but its value depends on whether detectors trained on simulated responses generalize beyond training targets. This work examines that question using acquisition-aware electromagnetic simulation of a D-band multiple-input multiple-output (MIMO) radar. Complex multistatic responses are generated before image formation, and simulated and measured data undergo the same coherent reconstruction and detector-input preparation. No measured target responses are used during development. Seven convolutional neural network (CNN) detectors are compared, with YOLOX-Tiny selected based on detection performance and inference cost. Across seven geometry-disjoint folds, YOLOX-Tiny achieves mean F2=0.89 and mAP0.5=0.88 for single-object scenes, and F2=0.82 and mAP0.5=0.79 for multi-object scenes. Generalization is further examined on 14 independent jointly simulated multi-object scenes and measurements from the reference D-band radar and two additional systems. The measurements provide qualitative evidence of transfer and reveal replica-related false detections and local class ambiguities in high-resolution responses. Lower-resolution measurements retain correctly detected targets alongside additional false detections assigned to other classes. These findings highlight the need for more representative synthetic radar responses and for relating image detections to physical targets.