Inpainting via high-dimensional universal shearlet systems (Q1782219)

This paper is on the asymptotic analysis of inpainting via high-dimensional universal shearlet systems. The paper is mainly based on the results of \textit{E. J. King} et al. [J. Math. Imaging Vis. 48, No. 2, 205--234 (2014; Zbl 1362.94007)] and \textit{M. Genzel} and \textit{G. Kutyniok} [SIAM J. Imaging Sci. 7, No. 4, 2301--2339 (2014; Zbl 1307.42033)]. First, the paper uses an abstract model for inpainting via \(\ell^1\) minimization, where a signal \(x^0\) is modeled as an element in a separable Hilbert space \(\mathcal{H} = \mathcal{H}_K\oplus\mathcal{H}_M\) with \(\mathcal{H}_K\) the known part space and \(\mathcal{H}_M\) the missing part space. Given a Parseval frame \(\Phi\) for \(\mathcal{H}\) and its associated analysis operator \(T_\Phi\), the \(\ell^1\) minimization problem for recovering a signal \(x^0\) with missing information is given by \[ x^* =\arg \min_{x\in \mathcal{H}} \|T_\Phi x\|_1, \text{ s.t. } P_K x = P_Kx^0. \] Using \(\delta\)-clustered sparsity and cluster coherence \(\mu_c(\Lambda, P_M\Phi)\) the reconstruction error of the signal \(x^*\) via the \(\ell^1\) minimization is controlled by the \(\delta\)-sparsity and the cluster coherence; see Theorem 1. The paper then introduces universal shearlet systems in high dimensions and shows that such systems can be built to be Parseval frames. Next, the paper introduces the image/video model using a distribution \(w\mathcal{L}\) in the Schwartz space and the missing mask by \(\mathcal M_h\). The remaining part of the section is then on choosing appropriate index sets for the shearlet coefficients so that both the \(\delta\)-clustered sparsity and cluster coherence \(\mu_c(\Lambda, P_M\Phi)\) are asympotically small. The main result of the paper is then stated in Theorem 3 that gives the asymptotic reconstruction error for the \(\ell^1\) minimization using the high-dimensional universal shearlet inpainting. Special case for dimension 2 is discussed and numerical experiments are conducted in the last part of the paper.