Regularization for Compressive Sensing
Shengcai Liu,Jiangshe Zhang,Junmin Liu,Qingyan Yin
2014-01-01
Abstract:Recently, the design of group sparse regularization has drawn much attentions in group sparse signal recovery problem. Two of the most popular group sparsity inducing regularization are the l1,2 and l1,∞ regularization, defined as the sum of l2 and l∞ norms respectively. Nevertheless, they may fail to simultaneously consider the intra-group and intergroup sparsity characteristic of the signal. For example, Huang and Zhang [14] assert that the l1,2 regularization is superior to the l1 regularization only for strongly group-sparse signals. This means the sparsity of intra-group is useless for l1,2 regularization. Our experiments show that recovering signals with intra-group sparse needs more measurements than those without by the l1,∞ regularization. In this paper, we propose a novel group sparsityinducing regularization defined as a mixture of the l1 norm and the l1/2 quasi-norm, referred to as l1,1/2 regularization, which can overcome the above shortcomings of l1,2 and l1,∞ regularization. We define a new null space property for l1,1/2 regularization and apply it to establish a recoverability theory for the both intra-group and inter-group sparse signals. In addition, we introduce an iteratively reweighted algorithm to solve this model, and analyze its convergence. Comprehensive experiments on simulated data show that our proposed l1,1/2 regularization is superior to l1,2 and l1,∞ regularization, especially for the signals with sparsity level low in inter-group and high in intra-group.