Plug-and-Play Synthetic Aperture Radar Image Formation Using Deep Priors

The reconstruction of synthetic aperture radar (SAR) images from phase history data is an ill-posed inverse problem which, in several lines of recent work, is solved by minimizing a cost function. Existing reconstruction methods use regularization to tackle the ill-posed nature of the imaging task. However, in general, these regularizers are either too simple to capture complex spatial patterns and can only promote fixed, predefined features, or lead to non-quadratic cost functions that are nontrivial to minimize. Recently emerging plug-and-play (PnP) priors technique is a flexible framework that allows forward models of imaging systems to be integrated with state-of-the-art regularizers. Inspired by this, we propose a novel PnP SAR image reconstruction framework for spotlight-mode SAR. SAR involves complex-valued reflectivities with spatial structure on the reflectivity magnitudes that can be learned and imposed as priors. This distinguishing aspect is formulated into our proposed framework. We demonstrate the use and effectiveness of a convolutional neural network (CNN) based prior model for the reconstruction of synthetic and real SAR scenes and compare the results with FFT-based, non-quadratic regularization-based, and dictionary learning-based reconstruction methods as well as a PnP framework with BM3D regularizer. Our results suggest that these deep priors enable the learning and incorporation of complicated spatial patterns more effectively than existing methods, and produce significantly improved images especially from limited observations.