Abstract
Greedy algorithms based on matching pursuit have received a lot of attention in recent times because of their ability to produce accurate, sparse, high-resolution representation of signals even when the input data are aliased. However, given the heuristic nature of the algorithms of this family, similar performances in terms of accuracy and quality of the signal representation come with very different computational costs and runtimes. We compare two matching pursuit alternatives in the context of the computation of 3D linear Radon transform: the basic and the orthogonal matching pursuit algorithms. We compare the two approaches in terms of quality of the signal representation, sparsity in the transformed domain, and runtime. Despite the greater computational cost involved in solving a least-squares optimization for all coefficients in orthogonal matching pursuit, we observe faster convergence, which is due to a smaller number of matching pursuit iterations. The higher quality of the estimated coefficients leads to higher accuracy in signal representation, i.e. smaller final residual. Our results are consistent with previous comparisons of the two approaches already presented in the literature for the case of signal reconstruction from random measurements. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 8:30 AM Presentation Start Time: 10:35 AM Location: 304B Presentation Type: Oral