Families of Tuned Scale-Space Kernels

Abstract

We propose a formalism for deriving parametrised ensembles of local neighbourhood operators on the basis of a complete family of scale-space kernels, which are apt for the measurement of a specific physical observable. The parameters are introduced in order to associate a continuum of a priori equivalent kernels with each scale-space kernel, each of which is tuned to a particular parameter value. Ensemble averages, or other functional operations in parameter space, may provide robust information about the physical observable of interest. The approach gives a possible handle on incorporating multi-valuedness (transparancy) and visual coherence into a single model. We consider the case of velocity tuning to illustrate the method. The emphasis, however, is on the formalism, which is more generally applicable.