Physics-Guided Surface Current Reconstruction of 3D PEC Targets from a Single Image via Cascaded Neural Networks

Accurate reconstruction of electromagnetic surface currents is essential for radar signature prediction, scattering-mechanism interpretation, and electromagnetic scene understanding. However, traditional computational electromagnetic methods require complete geometric models and fine surface meshing, resulting in high computational costs and limiting applicability in vision-based scenarios where only single-view images are available. This paper proposes a physics-guided cascaded neural network to infer the surface-current distribution of a target directly from a single optical image. The framework consists of three subnetworks for surface-normal estimation, depth reconstruction, and current refinement, together with an intermediate physics-guided module for coarse current initialization based on physical optics (PO). By progressively embedding geometric cues and electromagnetic priors, the proposed method effectively alleviates the ill-posed geometry-to-current mapping and enforces physical consistency. Extensive numerical experiments demonstrate that the proposed model accurately reconstructs high-fidelity surface currents and radar cross sections (RCS) at 1GHz, achieving agreement comparable to the method of moments (MoM) results while requiring only a single image as input. Experimental validations using 3D-printed objects further confirm the robustness of the framework under real-world measurement uncertainties, including illumination variations and sensor noise. The results show that the proposed method eliminates the need for full geometric modeling and offers strong potential for real-time radar signature prediction, EM simulation acceleration, and autonomous perception.