Parametric inversion of experimental electromagnetic imaging data using synthetically-trained neural networks
MetadataShow full item record
Electromagnetic imaging is a method that aims to reconstruct a spatial map of the material properties of a target through non-invasive measurements, allowing an observer to see inside objects, such as the human body, geological formations or large containers, without disturbing or entering them. This thesis evaluates the use of synthetically trained neural networks for performing phaseless parametric inversion of experimental electromagnetic imaging data. Fully-connected neural networks are used as a tool to obtain bulk parameters describing the geometry and complex-valued permittivity of materials within the region of interest (ROI) from experimental data. This data can subsequently be used for calibration and as prior information to facilitate full-phase 3D inversion of the data. These networks are applied to two different applications, one in agriculture and one in biomedical imaging, demonstrating that this method can be generalized to imaging systems at significantly different scales. In the agricultural application, synthetically-trained neural networks are used to infer bulk parameters from raw, uncalibrated experimental data for two different grain storage bins. This presents a promising tool for calibrating data collected from uncooperative systems where conventional calibration methods are not practical. For commercial grain monitoring systems, where multiple measurements are taken daily, neural networks present a cost-effective long-term alternative to existing iterative parametric inversion methods. For the biomedical application, a two-stage workflow is presented, demonstrating the suitability of the synthetically trained machine-learning-based parametric inversion to obtain the prior information needed for a second stage inversion using the Contrast Source Inversion (CSI) method. Noisy synthetic data representing a model of a human breast is used to simulate experimental data in the two-stage workflow, and preliminary bulk parametric inversion (neural network) results are given for experimental microwave breast imaging data. The results presented herein demonstrate that synthetically-trained neural networks can successfully infer bulk parameters from noisy synthetic and experimental data. Synthetic experiments for the microwave breast imaging case show that the inferred parameters are of sufficient quality to be used as prior information for full phase inversion and 3D image reconstruction.