Deep Learning Embedded Variable Pattern-Coupled SBL for Microwave Staring Correlated Imaging

Microwave staring correlated imaging (MSCI) is a novel imaging technique based on temporal-spatial stochastic radiation field, which can realize high resolution imaging. To further improve the performance of sparse recovery algorithm under block sparse scenarios, two-dimensional pattern-coupled sparse Bayesian learning (PCSBL-2D) and two-dimensional pattern-coupled sparse Bayesian learning with variable coupling coefficients (
$2\mathrm{D}$
-VPCSBL) have been proposed. However, the coupling prior of hyperparameters makes it impossible to obtain an analytic updating rule, thus a preset sub-optimal solution is adopted as an alternative during iterations, resulting a situation where hyperparameters are trapped into local minima. Therefore, a new deep learning embedded pattern-coupled sparse Bayesian learning network for MSCI is proposed, namely VPCSBL-Net. Based on the conventional
$2\mathrm{D}$
- VPCSBL algorithm framework, deep unfolding technique is adopted to unroll the iterations into a cascaded network, where the preset sub-optimal hyperparameters are set as learnable parameters, and the proposed network is trained via the back propagation (BP) algorithm. Experiments validate the superiority of the proposed network over 2D- VPCSBL algorithm from the perspective of convergence speed and performance under low signal-to-noise ratio (SNR) cases.