A deep iterative neural network for structured compressed sensing based on generalized pattern-coupled sparse Bayesian learning

Over the last few decades, lots of model-driven methods have been developed for model-based compressed sensing (CS), in which the inherent structures of sparse signals are exploited as priors to promote reconstruction accuracy and robustness. Although these methods obtained positive results, there still exist low prior utilizations. The single model-driven fashion is hard to completely express structured priors. Motivated by the recent flourishing studies of deep learning, we combine the model-based framework of the pattern-coupled sparse Bayesian learning (PC-SBL) with data-driven deep learning method to form a double-driven architecture, dubbed as PC-DINN. First, a generalized model of pattern -coupled Bayesian priors is developed to characterize structured properties, in which learnable scale parameters are generated by a heterogeneous process. Second, we unroll the iterative sparse Bayesian learning (SBL) algorithm to form an interpretable deep iterative neural network, and then treat all the learnable scale parameters of the prior model as weights to be learned. To the best of our knowledge, this is the first combination of model-based SBL and data-driven methods for structured CS. Simulation results suggest that for both the structured sparse signals of block sparsity and tree sparsity, the proposed PC-DINN not only achieves favorable reconstruction accuracy but also overcomes the vulnerability of parameter choice in the framework of PC-SBL.(c) 2022 Elsevier Inc. All rights reserved.