کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Blind Through-Wall Imaging via Differentiable Physics-AI Co-Design

نویسندگان :

Shokrollah Karimian¹ Mahdi Akhavan² Peyman Nasehpour³

1- دانشگاه شهید بهشتی تهران 2- دانشگاه شهید بهشتی تهران 3- آکادمی علوم نیویورک

کلمات کلیدی :

Differentiable Physics،Metasurface Antennas،Inverse Scattering،Computational Imaging

چکیده :

Through-Wall Imaging (TWI) stands at the intersection of electromagnetic theory and inverse problems, plagued by the “wall” dilemma: accurate image formation mandates precise knowledge of the wall’s constitutive parameters, yet these are rarely accessible in non-cooperative scenarios. Conventional analytical methods fail due to refractive defocusing, while emerging deep learning approaches suffer from poor generalizability and hallucination, treating the physical world as a black box. This paper introduces a paradigm shift towards Differentiable Physics-AI Co-Design. We effectively turn the electromagnetic wave equation into a learnable layer within a neural network. By differentiating the Green’s function propagation model with respect to the wall permittivity, we enable a single-port Meta surface Antenna to learn the physical environment blindly. This is achieved via a gradient descent loop that maximizes image contrast without requiring ground-truth training data. A secondary physics aware U-Net refines the output, strictly conditioning its generative capabilities on the physics-based reconstruction. Numerical results demonstrate that the proposed framework recovers wall permittivity with < 10% error and achieves high- fidelity shape recovery of metallic targets from highly compressed measurements, bridging the gap between rigorous physics and data-driven intelligence.