Combining Recurrent and Convolution structures: A Deep Compressed Sensing Network for Sparse Signal Recovery

Zhongjun Liu,Jun Zhang,Luhua Wang

DOI: https://doi.org/10.1145/3579654.3579718

2022-12-23

Abstract:With the powerful learning ability of neural networks, compressive sensing methods based on deep neural networks have shown great potential. However, most existing methods treat the neural network as a black box, which weakens the regularization of "known" signal priors. In order to solve this problem, a class of compressed sensing methods based on deep unfolding has been proposed in recent years. By mapping the iterative optimization algorithm in the classical model-based methods onto networks, the method based on deep unfolding integrates the advantages of both, making the network interpretable while greatly reducing the time complexity. In this paper, to solve the problem of sequential sparse reconstruction, we propose a novel model based on deep unfolding, dubbed DW-FISTA. The whole model can be divided into two modules, namely the mapping module and the refine module. The first module maps the iterative process of the fast iterative shrinkage-thresholding algorithm (FISTA) to an interpretable recurrent neural network composed of fixed phases. Compared with other ISTA models based on deep unfolding, our model has a better global convergence rate. In the second module, we use convolution filters and nonlinear activation functions to refine the reconstructed signal by taking the immediate reconstruction results of the mapping module as input. Experimental result shows that the proposed DW-FISTA model outperforms existing state-of-art models in sparse sequence reconstruction and can ensure a faster convergence rate.

Computer Science