Non-iterative microwave imaging solutions for inverse problems using deep learning

Abstract

This paper describes a U-net based Deep Learning (DL) approach in combination with Subspace-Based Variational Born Iterative Method (SVBIM) to provide a solution for the quantitative reconstruction of scatterer from the measured scattered field. The proposed technique can be used as an alternative to conventional time consuming and computationally complex iterative methods. This technique comprises a numerical solver (SVBIM) for generating the initial contrast function and a DL network to reconstruct the scatterer profile from the initial contrast function. Further, the proposed technique is validated against theoretical and experimental results available from the literature. Root Mean Square Error (RMSE) value is used as the metric to measure the accuracy of the reconstructed image. The RMSE values of the proposed method show a significant reduction in the reconstruction error compared with the recent Back Propagation-Direct Sampling Method (BP-DSM). The proposed method produces an RMSE value of 0.0813 against 0.1070 in the case of simulation (Austria Profile). The error value obtained by validating against the FoamDielExt experimental database in the case of the proposed method is 0.1037 against 0.1631 reported for BP-DSM method.