Fractional total variation denoising model with \(L^1\) fidelity (Q2145625)

Given \(f\in L^1(\mathbb{R}^N)\), and a parameter \(\Lambda>0\), the total variation-based image denoising model, with a \(L^1(\mathbb{R}^N)\) fidelity term, consists in finding functions attaining \[ \min\left\{ |D u| (\mathbb{R}^N) + \Lambda \int_{\mathbb{R}^N} |u-f|\, \mathrm{d}x\right\},\tag{1} \] among \(u \in BV(\mathbb{R}^N)\). The function \(f\) represents a given image (more precisely, its greyscale), which might be degraded by the presence of some noise, while any solution \(u\) represents a denoised image. Hence, physically, the problem corresponds to reconstruct images from some initial data. Problem 1. is a variant of the ROF model [\textit{L. I. Rudin} et al., Physica D 60, No.1--4, 259--268 (1992; Zbl 0780.49028)], where the fidelity term is given by the \(L^2\) norm in place of the \(L^1\) norm, granting convexity of the studied functional. It is worth noting that the loss of convexity is more than compensated by the higher sensitivity of the problem, yielding better numerical results. The author in this paper studies \[ \min\left\{\frac 12 \int_{\mathbb{R}^N}\int_{\mathbb{R}^N} \frac{|u(x)-u(y)|}{|x-y|^{N+s}}\, \mathrm{d}x\mathrm{d}y + \Lambda \int_{\mathbb{R}^N} |u-f|\, \mathrm{d}x\right\},\tag{2} \] where \(s\in (0,1)\), and the minimization runs in the fractional Sobolev space \(W^{s,1}(\mathbb{R}^N)\). It is the fractional counterpart of (1), since it is obtained by replacing the \(BV\) seminorm of \(u\) with the fractional \(W^{s,1}\) seminorm. This is motivated by the fact that the fractional order derivative term not only mantains the contour feature in the smooth area of the image, but also preserves high-frequency components like edges and textures. The author extends most of the results know for (1) to its fractional counterpart, see Theorems 1.2--1.7, also linking the problem to the fractional Cheeger problem. A key ingredient, which is of independent interest, is Theorem 1.1 that extends to the fractional setting a classic result by De Giorgi, establishing a uniform estimate of the difference between the \(s\)-fractional perimeter of a \(\mathrm{C}^{1,1}\) bounded set and that of any measurable set, w.r.t. their \(L^1\) distance.