Large-scale structured sparse image reconstruction with correlated multiple-measurement vectors using Bayesian learning

This paper proposes a Bayesian learning approach to structured sparse image reconstruction. In contrast to conventional paradigms which convert images into high-dimensional vectors and thus are impractical for recovering large-scale images, we formulate columns of image matrices into a multiple-measurement-vector (MMV) model to reduce the problem dimension. Besides, we simultaneously exploit the tree structure of image wavelet coefficients and the column correlations of image matrices in wavelet domain as two prior structured constraints to improve reconstruction accuracy. Experimental results reveal that our method significantly outperforms other MMV-based strategies in terms of reconstruction error and provides a practical and efficient alternative to large-scale structured sparse image reconstruction.