DAMAS on one million points: Greedy coordinate descent for the Covariance Matrix Fitting method for acoustical source imaging
Résumé
A variant of the DAMAS algorithm for acoustical source imaging by deconvolution is proposed, where the coefficient of the power map to be updated is chosen at each iteration, instead of being fixed in advance, or random. As the original DAMAS algorithm, the proposed algorithm converges to a solution of the Covariance Matrix Fitting problem. Experimental evaluation of the algorithm shows faster convergence than the DAMAS algorithm, with reduced memory usage (respectively 40 times faster, and 30 more efficient in memory in a 2D setting). The performances of the proposed method lie between CLEAN-SC, which performances are limited by its resolution, and the more performant but computationally intensive Lawson–Hanson algorithm and gridless method Sliding Frank-Wolfe. The computation complexity in space and time of the proposed method, both linear in function of the size of the discretization and number of microphones, allows imaging on a three dimensional grid of more than one million points, yielding sparser source distributions than with coarser grids in the case of point sources.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |