High-Speed Target HRRP Reconstruction Based on Fast Mean-Field Sparse Bayesian Unrolled Network

The rapid and accurate reconstruction of the high-resolution range profiles (HRRPs) of high-speed targets from incomplete wideband radar echoes is a critical component in space target recognition tasks (STRTs). However, state-of-the-art HRRP reconstruction algorithms based on sparse Bayesian learning (SBL) are computationally expensive and require the manual selection of prior scale parameters. To address these challenges, this paper proposes a model-driven deep network based on fast mean-field SBL (FMFSBL-Net) for the HRRP reconstruction of high-speed targets under missing data conditions. Specifically, we integrate a precise velocity compensation and HRRP reconstruction into the mean-field SBL framework, which introduces a unified SBL objective function and a mean-field variational family to avoid matrix inversion operations. To reduce the performance loss caused by mismatched prior scale parameters, we unfold the limited FMFSBL iterative process into a deep network, learning the optimal global prior scale parameters through training. Additionally, we introduce a sparsity-enhanced loss function to improve the quality and noise robustness of HRRPs. In addition, simulation and measurement experimental results show that the proposed FMFSBL-Net has a superior reconstruction performance and computational efficiency compared to FMFSBL and existing state-of-the-art SBL framework type algorithms.