On the stability of medial axis transform (Q874365)

Given a domain \(\Omega\subset {\mathbb{R}}\) the set \[ \text{MAT}(\Omega)=\{(p,r)\in {\mathbb{R}}^2\times [0,\infty)| B_r(p)\text{ is a maximal ball in }\Omega\}, \] is the medial axis transform of \(\Omega(\text{MAT}(\Omega))\), frequently used in shape analysis to reconstruct the original domain. The set \[ \text{MA}(\Omega)= \{p\in\mathbb R^2\mid \exists r\geq 0\text{ such that }(p,r)\in \text{MAT}(\Omega)\} \] is the medial axis of \(\Omega\). This paper extends a result by \textit{S. W. Choi} and \textit{H.-P. Seidel} [J. Math. Imaging Vis. 17, No. 3, 237--247 (2002; Zbl 1028.68186)] to general domains with natural boundary regularity. Let \({\mathcal{H}}(A| B)=\max _{p\in A}d(p,B)\). The main result proves one-sided stability for general domains: Theorem 1: If \(\Omega\) is a normal domain which is not weakly injective, then \[ {\mathcal{H}}\big(\text{MA}(\Omega)| \text{MA}(\Omega^\prime)\big)\leq K_\Omega \cdot \varepsilon ^{\frac{N_\Omega-1}{N_\Omega+1}}+o\left( \varepsilon ^{\frac{N_\Omega-1}{N_\Omega+1}}\right) \] \[ {\mathcal{H}}\big(\text{MAT}(\Omega)| \text{MAT}(\Omega^\prime)\big)\leq \sqrt{2}K_\Omega \cdot \varepsilon ^{\frac{N_\Omega-1}{N_\Omega+1}}+o\left( \varepsilon ^{\frac{N_\Omega-1}{N_\Omega+1}}\right) \] for every normal domain \(\Omega^\prime\) such that \(\max\{{\mathcal{H}}(\Omega,\Omega^\prime), {\mathcal{H}}(\partial\Omega, \partial\Omega^\prime)\}\leq \varepsilon\).