Anisotropic diffusion processes in early vision

Abstract

Summary form only given. Images often contain information at a number of different scales of resolution, so that the definition and generation of a good scale space is a key step in early vision. A scale space in which object boundaries are respected and smoothing only takes place within these boundaries has been defined that avoids the inaccuracies introduced by the usual method of low-pass-filtering the image with Gaussian kernels. The new scale space is generated by solving a nonlinear diffusion differential equation forward in time (the scale parameter). The original image is used as the initial condition, and the conduction coefficient c(x, y, t) varies in space and scale as a function of the gradient of the variable of interest (e.g. the image brightness). The algorithms are based on comparing the local values of different diffusion processes running in parallel on the same image.