Multi-Modal Physically Constrained Network for 3D Reconstruction of Complex Occluded Scenes in Terahertz Imaging and Its Applications

Abstract

Terahertz (THz) holographic imaging is hindered by occlusion and lack of physical consistency in data-driven reconstruction. This paper proposes a physics-constrained deep learning framework embedding THz wave propagation laws (angular spectrum theory and Fresnel diffraction) into network design. A dynamic multi-modal dataset simulating realistic occlusion across biomedical, industrial, and security scenarios is constructed. MT-JointNet jointly reconstructs object sequences and enforces physical plausibility via diffraction-consistency loss and physics-guided attention.