Adaptive image processing: first order PDE constraint regularizers and a bilevel training scheme (Q2693736)

A new variational method of a functional in image processing is defined to obtain a ``clean'' image from a version distorted by noise or defects in the digital transmission system. The image domain is assumed to be represented by the square \(Q:=(-(1/2),(1/2))^2\) and the images are functions in \(L^2(Q,\mathbb{R})\). A ``clean'' image is considered to be one that minimizes a functional of the form \[ I(u):=\Vert u-u_\eta\Vert_{L^2(Q,\mathbb{R})}^2+R_\alpha(u), \] where \(u_\eta\) is the distorted image and \(R_\alpha\) is a regularizing functional with \(\alpha\) being a parameter to be determined. To determine this parameter for obtaining the optimal ``clean'' image, a two-level semi-supervised method has been proposed by other researchers. This paper modifies that proposal by achieving satisfactory results for a certain set of training images. Some of the objectives of this paper, according to its authors, are: 1) To propose a new class of operators for image processing, which they call ``the PDE-constrained total generalized variation operators''. 2) ``the study of a training scheme optimizing the trade-off between effective reconstruction and fine image-detail preservation''. To understand the full paper it is necessary to have a good background in certain areas of Analysis, which are not usual in the formation of a professional Mathematician. The paper notes that it is worth further research in this area of image processing. References are extensive and very up-to-date.