A Qualitative Deep Learning Method for Inverse Scattering Problems

In this paper, we propose a novel deep convolutional neural network (CNN) based qualitative learning method for solving the inverse scattering problem, which is notoriously difficult due to its highly nonlinearity and ill-posedness. The trained deep CNN accurately approximates the nonlinear mapping from the noisy far-field pattern (from measurements) to a disk that fits the location and size of the unknown scatterer. The used training data is derived from the simulated noisy-free far-field patterns of a large number of disks with different randomly generated centers and radii within the domain of interest. The reconstructed fitting disk is also very useful as a good initial guess for other established nonlinear optimization algorithms. Numerical results are presented to illustrate the promising reconstruction accuracy and efficiency of our proposed qualitative deep learning method.