Loading...
Deep Neural Networks for Electromagnetic Inverse Scattering Problems in Microwave Imaging
Maricar, Mohammed Farook
Maricar, Mohammed Farook
Date
2023-11
Author
Advisor
Type
Thesis
Degree
Citations
Altmetric:
Description
A Master of Science thesis in Electrical Engineering by Mohammed Farook Maricar entitled, “Deep Neural Networks for Electromagnetic Inverse Scattering Problems in Microwave Imaging”, submitted in November 2023. Thesis advisor is Dr. Amer Zakaria and thesis co-advisor is Dr. Nasser Qaddoumi. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Abstract
Microwave imaging, due to its non-invasive and non-destructive nature of detection, is of interest in various applications such as biomedical imaging, geophysical surveying, and non-destructive testing. The most common technique for microwave imaging is by solving an electromagnetic inverse scattering (EMIS) problem. The high nonlinearity and ill-posedness of these EMIS problems have led researchers to develop various inversion algorithms, such as the distorted Born iteration method (DBIM), the contrast source inversion (CSI) method, and the sub-space optimization method (SOM). Though these conventional nonlinear inversion algorithms provide acceptable results in many applications with moderate size and contrast, their computational costs are usually very high. In addition, for high-contrast targets, the performance is generally degraded. To tackle these issues, various deep-learning techniques have been proposed in recent years. In this thesis, the EMIS problem will be solved using deep neural networks (DNNs), with a focus on solving two-dimensional (2D) microwave imaging problems. Different DNN architectures are tested with three types of complex inputs: the measured scattered field 〖(Eˢ ⃑ᶜᵗₘₑₐₛ〗^; an input image obtained from backpropagating (BP) the measured scattered field; or the novel physics-incorporated input (ITER10) obtained from the output image from the tenth iteration of CSI. For the purpose of training, validation, and testing the networks, 10000 samples from the well-known MNIST dataset are used. In the initial phase, two encoder-decoder-based convolutional neural networks (CNNs) are designed and implemented; the Baseline-AE, and the deep convolutional encoder-decoder network (DCEDnet). These models were tested in the two different input scenarios: (Eˢ ⃑ᶜᵗₘₑₐₛ and BP. The BP-DCEDnet provided improved results than the CSI at a much lesser time when tested within the MNIST database. In the next phase, three additional networks namely the Unet-Lite, Unet and the Attention- Unet (ATTN-Unet) together with DCEDnet were trained with BP and tested with complex profiles under different noise levels and permittivity range. The BP-ATTN- Unet outperformed the rest. However, the real component still needed to improve in comparison with the CSI. Thereby, the ATTN-Unet was trained with the ITER10 input to produce more stronger results. The optimized ITER10-ATTN-Unet was then successfully tested with experimental data and the results obtained were as expected.