Machine learning techniques for computationalradar imaging

Computational imaging radar, such as Synthetic Aperture Radar (SAR), is a remotesensing technology capable of providing imagery under all weather conditions andduring both day and night. In SAR, a synthetic aperture is created through themotion of the mounted platform, which transmits a wideband chirp signal. Uponreception, the received signal from the synthetic aperture is computationally synthesizedto form a SAR image. Despite its advantages, SAR is a complex, coherentimaging system that produces complex-valued data. It has inherent limitationsdue to factors like restricted bandwidth and limited look angles, which can leadto speckle. Additionally, uncertainties in modeling the physics of SAR introducefurther complexity, leading to phase errors.In this dissertation, we demonstrate solutions for SAR image reconstruction usingmachine learning techniques. We introduce sparsity-driven SAR imaging using convolutionaldictionary-based representation. However, this method does not performas well as k-SVD-based dictionary representation, due to its inherent limitationsin representation, particularly the requirement for piecewise smooth filters and lowglobal mutual coherence.We have developed a framework for SAR image reconstruction using denoisers asprior-incorporating functions in a Plug-and-Play prior configuration. BM3D-baseddenoisers and CNN-based denoisers are explored, resulting in state-of-the-art performance.While CNNs as denoisers yield superior performance, they also have the potentialto remove artifacts and perform complex mapping from noisy to clean images. SARimage 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 ofcomplex-valued measurements.Although CNNs have demonstrated remarkable performance in SAR image reconstruction,they often lead to a washed-out or blurred effect. We hypothesize thatthis issue arises from the loss function. To address this, we designed a new lossfunction that balances denoising and texture preservation. The novel loss functionis based on the logarithmic discrete cosine transform, resulting in state-of-the-artperformance compared to other commonly used loss functions.The uncertainty in modeling SAR physical phenomena is evident, resulting in phasenoise and making SAR image reconstruction (autofocus) a significant challenge. Weintroduce a framework based on semi-supervised, CNN-driven autofocus, which notonly offers performance gains but also holds potential for further improvement inSAR image reconstru