Minimal surfaces: A geometric three dimensional segmentation approach (Q1374743)

The problem attacked is that of analytic representation of a given object whose characteristic function \(\chi\) is known; let I be a regularized version of \(1-\chi\). Then one starts with a surface, preferably given as a level surface of a function \(u(x,y,z)\) that encloses the desired object and constructs approximations to the flow \(\partial u/ \partial t= |\nabla u|\text{div} [g(I) |\nabla u|^{-1} \nabla u] +\nu g(I) |\nabla u|\), where \(g\) is a given weight function that vanishes for \(I=0\) and \(\nu\) is a positive constant. This can be considered as a minimal flow in some Riemannian metric. The constant \(\nu\) must be well chosen to provide an effective algorithm (if it is too large, the algorithm may fail to detect changes in the topology; if it is very small, the algorithm becomes very slow). The consideration of existence and uniqueness of the flow gives some indication on the choice of \(\nu\) for a given problem. Examples are given, including one where a topological sphere is deformed automatically to approximate two linked tori.