Loading...

Electromagnetic Imaging Using Machine Learning

Ahmadi, Leila | 2024

0 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 58169 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Shishegar, Amir Ahmad
  7. Abstract:
  8. This dissertation focuses on addressing electromagnetic forward and inverse scattering problems using physics-based deep learning models. A key challenge in integrating physics with data-driven models lies in the implicit nature and nonlinearity of the forward scattering operator. To address this, a novel closed-form expression is proposed for piecewise homogeneous media. By formulating the scattered field as an implicit function of the permittivity values in each homogeneous region, the computational complexity of solving the forward problem for a fixed geometry is changed from O(N^3 ) to O(NΣM_k^2), where N is the total number of meshes, and M_k refers to the meshes in the kth region. This significantly reduces complexity when dealing with only a few piecewise homogeneous regions. Nonlinearity is mitigated using two distinct iterative methods: Jacobi iteration method and the Born series. We combined both algorithms with deep learning models by unrolling them into a single framework. Although the Jacobi iteration-based model did not demonstrate satisfactory generalization capabilities, the Born series-based model, designated as DL-BA series, shows convergence for comparably stronger scatterers than Born approximation and achieves a mean squared error on the order of 〖10〗^(-3) when predicting the total field in free space. This series shows a very good generalization capability due to incorporating physics in its model design. The ill-posed nature of the inverse problem is also considered by exploring methods to integrate a priori information into data-driven models. Using a variational auto-encoder to map the image space (representing the solution space) to a latent space, we mapped both a priori information and the measurement data into this low-dimensional space. This method is particularly applicable when imaging specific body parts, demonstrating good noise robustness and the ability to support strong scatterers
  9. Keywords:
  10. Microwave Imaging ; Direct/Invers Scattering ; Deep Learning ; Machine Learning ; Inverse Scattering

 Digital Object List

 Bookmark

...see more