Abstract
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.
Original language | English (US) |
---|---|
Pages (from-to) | 153-160 |
Number of pages | 8 |
Journal | Applied Computational Electromagnetics Society Journal |
Volume | 35 |
Issue number | 2 |
State | Published - 2020 |
Keywords
- Convolutional neural network
- Deep learning
- Inverse acoustic scattering
- Qualitative method
ASJC Scopus subject areas
- Astronomy and Astrophysics
- Electrical and Electronic Engineering