An $\ell_p$ Variable Projection Method for Large-Scale Separable Nonlinear Inverse Problems

TL;DR

This paper introduces a modified variable projection method for large-scale separable nonlinear inverse problems, incorporating edge-preserving regularization and automatic parameter selection to improve image reconstruction quality.

Contribution

It proposes a novel majorization minimization approach that efficiently handles $ extit{ ext{l}}_p$ regularization, including TV, framelet, and wavelets, for large-scale inverse problems.

Findings

Enhanced image reconstruction quality demonstrated on blind deconvolution problems.

Automatic regularization parameter selection improves convergence and results.

Method efficiently solves large-scale problems with generalized Krylov subspace methods.

Abstract

The variable projection (VarPro) method is an efficient method to solve separable nonlinear least squares problems. In this paper, we propose a modified VarPro for large-scale separable nonlinear inverse problems that promotes edge-preserving and sparsity properties on the desired solution and enhances the convergence of the parameters that define the forward problem. We adopt a majorization minimization method that relies on constructing a quadratic tangent majorant to approximate a general $\ell_p$ regularized problem by an $\ell_2$ regularized problem that can be solved by the aid of generalized Krylov subspace methods at a relatively low cost compared to the original unprojected problem. In addition, we can use more potential general regularizers including total variation (TV), framelet, and wavelets operators. The regularization parameter can be defined automatically at each iteration by means of generalized cross validation. Numerical examples on large-scale two-dimensional imaging problems arising from blind deconvolution are used to highlight the performance of the proposed approach in both quality of the reconstructed image as well as the reconstructed forward operator.