Mean curvature diffusion is shown to be a position vector diffusion, t
ending to scalar diffusion as a flat image region is approached, and p
roviding noise removal by steepest descent surface minimization. At ed
ges, it switches to a nondiffusion state due to two factors: the Lapla
cian of position vanishes and the magnitude of the surface normal atta
ins a local maximum. (C) 1998 Elsevier Science B.V. All rights reserve
d.