Deep Learning for Joint Image Reconstruction and Segmentation for SAR

We present an approach for joint image reconstruction and foreground-background separation for synthetic aperture radar (SAR) using deep learning (DL). Network structure of the deep model is derived by unwrapping the stages of an iterative algorithm that solves an underlying optimization problem. This leads to physical model based deep network with learned network parameters having meaningful interpretation. Combined image reconstruction and segmentation approach allows joint optimization of both tasks that enhances performance and prevent inadvertent loss of useful information. Numerical results are included to show feasibility of the proposed approach.