A general fractional total variation-Gaussian (GFTG) prior for Bayesian inverse problems

TL;DR

This paper introduces a novel general fractional total variation-Gaussian (GFTG) prior for Bayesian inverse problems, improving image reconstruction by reducing staircase effects and enhancing textures, with theoretical analysis and numerical validation.

Contribution

The paper develops a new GFTG prior combining fractional derivatives with Gaussian regularization, extending prior models and providing theoretical and computational frameworks for infinite-dimensional Bayesian inverse problems.

Findings

GFTG prior reduces staircase artifacts in images.

Enhanced texture details compared to traditional priors.

Numerical examples demonstrate improved reconstruction quality.

Abstract

In this paper, we investigate the imaging inverse problem by employing an infinite-dimensional Bayesian inference method with a general fractional total variation-Gaussian (GFTG) prior. This novel hybrid prior is a development for the total variation-Gaussian (TG) prior and the non-local total variation-Gaussian (NLTG) prior, which is a combination of the Gaussian prior and a general fractional total variation regularization term, which contains a wide class of fractional derivative. Compared to the TG prior, the GFTG prior can effectively reduce the staircase effect, enhance the texture details of the images and also provide a complete theoretical analysis in the infinite-dimensional limit similarly to TG prior. The separability of the state space in Bayesian inference is essential for developments of probability and integration theory in infinite-dimensional setting, thus we first introduce the corresponding general fractional Sobolev space and prove that the space is a separable Banach space. Thereafter, we give the well-posedness and finite-dimensional approximation of the posterior measure of the Bayesian inverse problem based on the GFTG prior, and then the samples are extracted from the posterior distribution by using the preconditioned Crank-Nicolson (pCN) algorithm. Finally, we give several numerical examples of image reconstruction under liner and nonlinear models to illustrate the advantages of the proposed improved prior.