A Learning Bayesian MAP Framework for Joint SAR Imaging and Target Detection

Abstract

In synthetic aperture radar (SAR) information acquisition, target detection is often performed on the basis of the acquired radar images. Under low signal-to-clutter ratio (SCR) or low signal-to-noise ratio (SNR) conditions, detection by images is likely to cause loss of targets. To address this problem, we propose a joint imaging and target detection network based on Bayesian maximum a posteriori (MAP) estimation. The imaging and detection results are, respectively, defined as scene magnitude and detection label, and their joint probability distribution is used in place of the distribution of scene magnitudes. In the MAP estimation, the continuity feature of the detection label is merged into the optimization process, and the imaging and detection results are optimized alternately to get an iterative solution. The iterative solution is then unrolled into a network, which consists of three modules. We first utilize the unrolled fast iterative shrinkage thresholding algorithm (FISTA) method for the image formation module and then incorporate the detection label estimation module and distribution parameter updating module to learn the detection label and the function of distribution parameters. This approach applies prior information for both imaging and detection processes and enables automatic learning of parameters that are difficult to fit. Simulation experiments demonstrate that the method can simultaneously achieve imaging and target detection under strong clutter and strong noise conditions, showing superior performance in both aspects.