Weighted double-backtracking matching pursuit for block-sparse reconstruction

Abstract

This study presents a new method for the reconstruction of block-sparse signals with and without noisy perturbations, termed weighted double-backtracking matching pursuit (WDBMP). Unlike anterior block-sparse reconstruction algorithms, WDBMP requires no prior knowledge about block length and boundaries. It not only refines the current approximation based on energy, but also takes advantage of block structure to refine the chosen support set, and thus to improve the recovery performance. Moreover, the authors propose weighted proxy to select the candidates, which can increase the probability of selecting correct supports and improve the convergence speed. Experimental results show that the proposed algorithm owns better recovery quality and requires fewer iterations to converge compared with the existing block-sparse reconstruction algorithms without knowing the block-sparse boundaries.