DATA-DEPENDENT SAMPLING OF 2-DIMENSIONAL SIGNALS

Samples taken at scattered points of a finite-support two-dimensional signal can be interpolated to recover an approximation of the original signal. Given a bound on the number of samples, where should they be placed to enable the most accurate reconstruction? Or, given an error bound for the reconstruction, what is the minimum number of samples re quired, and where should they be placed? In this paper we introduce se arch schemes that provide good candidate solutions to these problems, for digital signals. Natural Neighbour Interpolation is used in iterat ive sample removal and movement processes to obtain sparse sample patt erns. For pictures and Digital Elevation Models, fewer samples are req uired if the interpolant is only C-0 continuous at the data sites, tha n if it is C-1. Retained samples lie on the ridges and valleys of the laplacian.