Optical Image-Based Forward Scattering Modeling Method via Physics-Embedded Deep Learning

In this paper, a physics-embedded deep learning method for forward scattering modeling of targets is proposed. It performs direct prediction of scattering centers and synthetic aperture radar (SAR) images using optical images taken from diverse viewpoints of the target. Comparing to traditional scattering modeling methods, it can avoid the complex and time-consuming steps, such as geometric modeling, electromagnetic simulation, SAR imaging, etc. The proposed method is composed of two modules. The first is an end-to-end supervised deep learning module designed to generate SAR images from input geometric images of target. The second is a zero-shot physical module which employs the segment anything model for semantic segmentation and uses the parametric electromagnetic part model to derive the scattering response of the target. By integrating prior knowledge from the physics module into the deep learning module, it is possible to reduce the amount of data samples and significantly enhance the generalization capability of the model. Numerical results are presented to validate the effectiveness of the proposed method.