Date

2022-6-29

Author

Manisalı, İrfan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

642
views

165
downloads

Cite This

Near-field multiple-input multiple-output (MIMO) radar imaging systems are of interest in diverse fields such as medicine, through-wall imaging, airport security, and surveillance. These computational imaging systems reconstruct the three-dimensional scene reflectivity distribution from the radar data. Hence their imaging performance largely depends on the image reconstruction method. The analytical reconstruction methods suffer from either low image quality or high computational cost. In fact, sparsity-based methods offer better image quality than the traditional direct inversion methods, but their high computational cost is undesirable in real-time applications. In this thesis, we develop two novel deep learning-based reconstruction methods for near-field MIMO radar imaging. The main goal is to achieve high image quality with low computational cost. The first approach has a two-staged structure that consists of an adjoint operation followed by a deep neural network. The adjoint stage exploits the observation model and back project the measurements to the reconstruction space. The second stage employs a deep neural network which is trained to convert the backprojected measurements to a suitable reflectivity image. For comparison, a second approach is also developed which replaces the adjoint stage with a fully connected neural network. In this two-staged structure, the reconstruction is performed directly from the radar measurements using neural networks which are trained end-to-end to learn the direct mapping between the measurements and unknown reflectivity magnitude. For each case, a 3D U-Net is used at the second stage to jointly exploit range and cross-range correlations. We demonstrate the performance of the developed methods using a synthetically generated dataset and compare with the commonly used analytical methods. The developed two-staged method with adjoint provides the best reconstruction quality while enabling fast reconstruction.

Subject Keywords

Computational imaging, Image restoration, Image reconstruction, Inverse problems, Deep learning, Convolutional neural networks, Multiple-input multiple-output radar imaging, Near-field microwave imaging

URI

https://hdl.handle.net/11511/98756

Collections

Graduate School of Natural and Applied Sciences, Thesis