The shell-based inpainting Gaussian-blur conjecture

From papers

Let the inpainting domain DD and undamaged area U\mathcal U, together with their discrete counterparts DhD_h and Uh\mathcal U_h, be as described in the paper. Let u0:URdu_0:\mathcal U\to\mathbb R^d be non-negative and bounded, and let uh:DhRdu_h:D_h\to\mathbb R^d be produced by Algorithm 1 or its semi-implicit extension. Write ara_r^* for the stencil, let Zi=(Vi,Wi)Z_i=(V_i,W_i) have support in Supp(ar)\operatorname{Supp}(a_r^*) and mean (μx,μy)(\mu_x,\mu_y), and let Πθr:DD\Pi_{\theta_r^*}:D\to\partial D be the transport operator with direction gr=(μx,μy)\mathbf g_r^*=(\mu_x,\mu_y). Define the discrete mollification by

(u0y=0gσ(h))(x):=i=1N[(i1)hihgσ(h)(xt),dt]u0y=0(ih),\left(u_0\big|_{y=0}*g_{\sigma(h)}\right)(x):=\sum_{i=1}^N\left[\int_{(i-1)h}^{ih}g_{\sigma(h)}(x-t)\\,dt\right]u_0\big|_{y=0}(ih),

where

σ(h)2=γ2yhμy3,γ2=Var(μxW1μyV1).\sigma(h)^2=\frac{\gamma^2yh}{|\mu_y|^3},\qquad \gamma^2=\operatorname{Var}(\mu_xW_1-\mu_yV_1).

The shell-based inpainting Gaussian-blur conjecture. If u0u_0 is independent of its yy-coordinate, then

uh(x,y)(u0y=0gσ(h))(Πθr(x,y))u_h(x,y)\to\left(u_0\big|_{y=0}*g_{\sigma(h)}\right)(\Pi_{\theta_r^*}(x,y))

asymptotically as h0h\to0, at a rate dependent on yy but independent of xx. For general u0u_0, the same conclusion holds with u0y=0u_0\big|_{y=0} replaced by u~0\tilde u_0, where u~0(ih)=j=r10αj(ih)u0(ih,jh)\tilde u_0(ih)=\sum_{j=-r-1}^{0}\alpha_j(ih)u_0(ih,jh), 0αj(ih)10\leq\alpha_j(ih)\leq1, and j=r10αj(ih)=1\sum_{j=-r-1}^{0}\alpha_j(ih)=1. The context is numerical evidence that shell-based geometric inpainting produces a blur described by a one-dimensional Gaussian whose variance depends on the distance travelled in the inpainting direction; the claim concerns the asymptotic behaviour of the algorithm as the grid spacing tends to zero.

Progress summary

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Sources & referencesView supporting material

Primary source

L. Robert Hocking, Thomas Holding and Carola-Bibiane Schoenlieb, “Numerical analysis of shell-based geometric image inpainting algorithms and their semi-implicit extension”, arXiv:1707.09713 (2017).

Solutions 0

No solutions have been posted yet.