Geometrically Guided Exemplar-Based Inpainting

Abstract

Exemplar-based methods have proven their efficiency for the reconstruction of missing parts in a digital image. Texture as well as local geometry are often very well restored. Some applications, however, require the ability to reconstruct non local geometric features, e.g. long edges. We propose in this paper to endow a particular instance of exemplar-based method with a geometric guide. The guide is obtained by a prior interpolation of a simplified version of the image using straight lines or Euler spirals. We derive from it an additional geometric penalization for the metric associated with the exemplar-based algorithm. We discuss the details of the method and show several examples of reconstruction.

Affine Invariant Self-similarity for Exemplar-based Inpainting

“…As pointed out in (Cao et al, 2011), adding transformations of patches makes it even more complex. Intuitively, the added variability makes it harder to distinguish "good" minima from other minima (a single pixel can be scaled to match a constant patch).…”

“…The geometry-oriented methods formulate the inpainting problem as a boundary value problem and the images are modeled as functions with some degree of smoothness expressed, for instance, in terms of the curvature of the level lines (Masnou and Morel, 1998;Ballester et al, 2001;Masnou, 2002;Chan and Shen, 2001b;Cao et al, 2011), with propagation PDE's (Bertalmío et al, 2000), or as the total variation of the image (Chan and Shen, 2001a). These methods perform well in propagating smooth level lines or gradients, but fail in the presence of texture or big inpainting domains.…”

“…The high dimensionality of the parameter space makes the search problem very computationally expensive and the excessive variability of candidates may lead to unstable results. In order to restrict the search space, authors of (Cao et al, 2011) propose to combine an exemplar-based approach that includes all rotated patches, with a geometric guide computed by minimizing Euler's elastica of contrasted level lines in the inpainted region.…”

Multiscale Analysis for Images on Riemannian Manifolds

“…This simple and fast approximative solution was enough to prevent the appearance of border effects in our results without significantly increasing the computational time involved in processing each frame. Nevertheless, note that this is just a practical choice for the sake of efficiency and that more accurate image extension methods can be used, for instance, image inpainting [23,14,5].…”

Tensor-Directed Spatial Patch Blending for Pattern-Based Inpainting Methods

“…2) The hybrid methods [16,8] that incorporate geometry-based methods to first continue the main structures before inpainting the textures. These methods are quite slow in practice, and require a segmentation of the image to separate structures from textures, which is an ill-posed problem.…”