Point source super-resolution via non-convex \(L_1\) based methods (Q333212): Difference between revisions
From MaRDI portal
Latest revision as of 19:37, 12 July 2024
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Point source super-resolution via non-convex \(L_1\) based methods |
scientific article |
Statements
Point source super-resolution via non-convex \(L_1\) based methods (English)
0 references
28 October 2016
0 references
In the super-resolution problem, one has to recover sparse signals \(x\in {\mathbb R}^N\) consisting only of separated peaks from frequency measurements up to a low frequency cutoff \(f_c\). Then \(\lambda_c = 1/f_c\) is called Rayleigh length of \(x\). If the peak separation \(\Delta\) of \(x\) fulfills \(\Delta \geq 2\lambda_c N\), then such sparse signal \(x\) can be recovered by \(\ell_1\) minimization (see [\textit{E. J. Candès} and \textit{C. Fernandez-Granda}, Commun. Pure Appl. Math. 67, No. 6, 906--956 (2014; Zbl 1350.94011)]). In this paper, the authors analyze the so-called minimum separation factor MSF = \(\Delta/(\lambda_c\, N)\). Thus MSF \(\geq 2\) guarantees the exact recovery by \(\ell_1\) minimization. Numerical tests show that \(\ell_1\) minimization often fails when MSF \(< 1\). In this case, the authors investigate the recovery of \(x =(x_j)_{j=1}^N\) by non-convex \(\ell_1\) based minimization with the regularization term \(R(x) = \| x\|_1 - \| x\|_2\) or \[ R(x) = \sum_{j=1}^N \min \{|x_j|, \,\alpha\} \] with some \(\alpha>0\). The non-convex \(\ell_1\) based minimization problem is solved via a difference of convex algorithms such that local minimizers are obtained. The authors show several properties of the local minimizers. Numerical experiments for signals and images are examined.
0 references
super-resolution
0 references
Rayleigh length
0 references
minimum separation factor
0 references
non-convex \(\ell_1\) based minimization
0 references
reconstruction of sparse signals
0 references
\(\ell_1\) minimization
0 references
numerical test
0 references
regularization
0 references
numerical experiment
0 references
0 references
0 references
0 references