PhD.Dissertation Defense:Ammar Saleem

MACHINE LEARNING TECHNIQUES FOR COMPUTATIONAL RADAR IMAGING

Ammar Saleem
Electronics Engineering, PhD Dissertation, 2024

Thesis Jury

Prof. Selim Balcısoy (Dissertation Supervisor), Prof. Özgür Gürbüz, Assoc. Prof. Erchan Aptoula, Asst. Prof. Lavdie Rada Ülgen , Asst. Prof. Muhammed Burak Alver

Prof. Müjdat Çetin (Dissertation Co-Supervisor)

Date & Time: December 16, 2024,  4.30 - 6.30 PM

Place: https://sabanciuniv.zoom.us/my/balcisoy

Keywords : Synthetic aperture radar, inverse problems, computational imaging,

deep learning, convolutional neural networks, plug-and-play priors, Discrete Cosine

Transform, loss functions

Abstract

Computational imaging radar, such as Synthetic Aperture Radar (SAR), is a remote sensing technology capable of providing imagery under all weather conditions and during both day and night. In SAR, a synthetic aperture is created through the motion of the mounted platform, which transmits a wideband chirp signal. Upon reception, the received signal from the synthetic aperture is computationally synthesized to form a SAR image. Despite its advantages, SAR is a complex, coherent imaging system that produces complex-valued data. It has inherent limitations due to factors like restricted bandwidth and limited look angles, which can lead to speckle. Additionally, uncertainties in modeling the physics of SAR introduce further complexity, leading to phase errors.

In this dissertation, we demonstrate solutions for SAR image reconstruction using machine learning techniques. We introduce sparsity-driven SAR imaging using convolutional dictionary-based representation. We have developed a framework for SAR image reconstruction using denoisers as prior-incorporating functions in a Plug-and-Play prior configuration. While CNNs as denoisers yield superior performance, they also have the potential to remove artifacts and perform complex mapping from noisy to clean images. SAR image reconstruction using full capability of a CNN was therefore required. We developed a framework that uses a generative network for SAR image reconstruction, incorporating a prior-induced loss function in addition to CNN-based denoising of complex-valued measurements.

Although CNNs have demonstrated remarkable performance in SAR image reconstruction, they often lead to a washed-out or blurred effect. We hypothesize that this issue arises from the loss function. To address this, we designed a new loss function that balances denoising and texture preservation. The novel loss function is based on the logarithmic discrete cosine transform, resulting in state-of-the-art performance compared to other commonly used loss functions.

The uncertainty in modeling SAR physical phenomena is evident, resulting in phase noise and making SAR image reconstruction (autofocus) a significant challenge. We introduce a framework based on semi-supervised, CNN-driven autofocus, which not only offers performance gains but also holds potential for further improvement in SAR image reconstruction under model uncertainty.