Abstract
Electromagnetic inverse scattering imaging (EISI) is a highly challenging problem due to its nonlinear and ill-posed nature. Traditional optimization methods suffer from high computational complexity and limited robustness, while deep learning approaches are criticized for their lack of interpretability and restricted generalization capabilities. To address these challenges, we propose a kernelized deep unfolding network framework with joint physical constraints, in which a kernelized U-Net module is employed at each stage to predict the induced currents, while the scattered field and permittivity are explicitly computed using physical operators. All these physical quantities are incorporated into the loss function, thereby enforcing strict compliance with electromagnetic laws throughout the learning process. Furthermore, at the bottleneck of the kernelized U-Net, a kernelized convolution module and a random Fourier features module are integrated—the former captures local nonlinear spatial features, including sub-wavelength structures, while the latter extracts global spectral information to characterize wavelength-scale interference patterns. Extensive evaluations on both numerical simulations and experimentally measured data demonstrate that the proposed method consistently outperforms existing techniques in reconstruction accuracy, generalization capability, and physical consistency, highlighting its robustness and practical potential for EISI.
Funding Information
  • National Natural Science Foundation of China (62027803)
  • National Natural Science Foundation of China (61971111)
  • National Natural Science Foundation of China (61801089)
  • National Natural Science Foundation of China (61601096)
  • National Natural Science Foundation of China (6170109)

This publication has 59 references indexed in Scilit: