A Multilayer Joint Regularized Block Sparse Bayesian Algorithm With Space-Time Structured Prior Learning Function

The structured sparse model incorporates dependencies among coefficients, and exploiting intrablock correlations significantly enhances recovery performance. The reconstruction performance of existing block sparse recovery algorithms relies on prior knowledge of blockwise sparse coefficients, including the number of blocks, boundaries, and sizes, which severely restricts the applicability of algorithms in scenarios like radar reconstruction. This article proposes a regularized block sparse Bayesian algorithm with structured prior learning, which involves a two-stage processing procedure. The first stage includes the acquisition of space-temporal hierarchical prior information. It introduces a regularization penalty term in the cost function to constrain and optimize hyperparameters, thereby promoting block correlations among neighboring coefficients. In the second stage, refined recovery is executed for nonzero blocks using a model with multiple measurement vectors (MMVs) constructed from space-temporal dimensional data. This model captures temporal correlations among multiple snapshots and blocks correlations within the prior coefficient structure, enabling joint parameter reconstruction across multiple domains. Simulation results demonstrate that compared with the conventional block sparse Bayesian learning (BSBL) algorithm, the proposed approach exhibits significant advantages in reconstruction performance under an unknown block sparse model.