Unsupervised Deep Learning Parameter Estimation for High Fidelity Synthetic Aperture Radar Super Resolution

Super-Resolution for Synthetic Aperture Radar (SAR) has been of high interest for its applications in reducing system Size Weight and Power (SWAP) and potential to improve SAR interpretability. Yet to enable super-resolution techniques like Super Spatially Variant Apodization (SSVA), precise knowledge of the Synthetic Aperture Radar (SAR) imaging parameters is required. Such parameters include the pixel, ground range and azimuth resolution which may be missing or tedious to parse. Moreover, these parameters are dependent on terrain relief and the effective target beam-width which may be absent in practical scenarios involving real targets. The inability to calculate sampling parameters would degrade side-lobe levels and image quality in super-resolution processing. To tackle this, we design a novel deep learning network that leverages a bi-linear sampling layer and a total variation loss that is able to directly estimate the required sampling factor without supervision. Our network is able to learn and carry out optimal side-lobe cancellation with no prior knowledge of the system or target imaging parameters. In doing so, we make it possible to carry out super-resolution on SAR images with no a-prior expert knowledge and outperform traditional algorithms in presence of imperfect expert knowledge. We validate our approach on both simulated data and a dataset (GOTCHA) collected by Airforce Research Laboratory (AFRL).