An Accelerated MMP With a Pruning Tree Strategy

Abstract

Multipath matching pursuit (MMP) has been developed to solve the sparse signal recovery problems in compressed sensing, which can generate recovery error less than the traditional orthogonal matching pursuit type algorithms in terms of mean square error. However, the computational burden of MMP is seriously heavy, and limits itself in applications, finding the need to be improved urgently. To lighten the burden, in this paper, an accelerated version of the MMP is proposed based on pruning tree strategy for sparse signal recovery, that attempts to reduce some of the paths in subsequent searching. By this strategy, an accelerated MMP is established in the scenarios where the MMP works. It retains the desirable performance as MMP and can shorten the running time. We demonstrate the acceleration ability of the proposed algorithm with almost the same accuracy of the original MMP, for the reconstruction of both synthetic data and images. INDEX TERMS Compressed sensing, multipath matching pursuit, sparse recovery, pruning tree strategy, restricted isometry constant.

A robust matching pursuit algorithm using information theoretic learning

“…Tree search and neural network based matching pursuit is another important group which mainly focuses on learning deep features from multiple paths [14][15][16] or from single hidden neural network [17]. In [14], the authors proposed a multipath hierarchical matching pursuit to learn features by capturing multiple aspects of discriminative structures of the data in a deep path architecture.…”

“…In [14], the authors proposed a multipath hierarchical matching pursuit to learn features by capturing multiple aspects of discriminative structures of the data in a deep path architecture. Algorithms in [15] and [16] are tree search based methods which use different deep tree search strategies during feature selection and estimation procedures to improve the sparse approximation. In [17], the authors proposed a learned OMP algorithm for fast sparse approximation by learning a single hidden neural network for subspace clustering.…”