Neural network enhancement of the Born approximation
We examine the use of neural networks in conjunction with the Born approximation for solving the inverse scattering problem for a penetrable medium in the frequency domain. The goal is a fast method for determining the refractive index of a scatterer from multistatic far-field measurements at one frequency. We present a new neural network model that seeks to improve on our previous results in by sharpening the reconstructed images. In addition, we present two other networks motivated by the corrected Born method that avoid the need to invert the regularized Born operator. These latter algorithms will be important for future three dimensional applications when the inverse Born approximation is difficult to compute, or when near-field measurements are used. Computational results illustrate that the new corrections to the Born approximation can provide an accurate and flexible approach to the inverse scattering problem.