Physics-Informed 2-D Convolutional Vision Transformer for Super-Resolution Microwave Coincidence Imaging

In this paper, a physics-informed Two-Dimensional (2-D) Convolutional Vision Transformer (ConvViT-2D) network is proposed to achieve super-resolution Microwave Coincidence Imaging (MCI). First, a modified Gram-Schmidt orthogonalization process is implemented to extract amplitude and phase information from detected echo signals. Subsequently, the ConvViT-2D network, featuring a multi-scale interference-aware tokenizer and a dimension-collapse pooling mechanism, is developed to estimate target positions. The ConvViT-2D-predicted support sets are then incorporated into a standard Orthogonal Matching Pursuit (OMP) residual update, which removes the contributions of previously identified targets and provides progressively cleaner residual features for detecting weaker scattering targets. Finally, a Tikhonov-regularized least-squares inversion is performed to retrieve the scattering coefficients of the imaging scene. Both simulated and experimental results demonstrate a tenfold improvement in MCI resolution, enabling the discrimination of closely spaced targets separated by one-tenth of the antenna’s 3-dB beamwidth.

Citation: P. Li#, M. Zhao#, I. Yoo, T. Fromentèze, M. Zhang, S. Zhu, and O. Yurduseven, "Physics-informed 2-D convolutional vision transformer for super-resolution microwave coincidence imaging," IEEE Trans. Radar Syst., early access, Sept. 2026.

Back