We introduce a nonlinear refinement subdivision scheme based on median-inte
rpolation. The scheme constructs a polynomial interpolating adjacent block
medians of an underlying object. The interpolating polynomial is then used
to impute block medians at the next ner triadic scale. Perhaps surprisingly
, expressions for the refinement operator can be obtained in closed-form fo
r the scheme interpolating by polynomials of degree D = 2. Despite the nonl
inearity of this scheme, convergence and regularity can be established usin
g techniques reminiscent of those developed in analysis of linear refinemen
t schemes.
The refinement scheme can be deployed in multiresolution fashion to constru
ct a nonlinear pyramid and an associated forward and inverse transform. In
this paper we discuss the basic properties of these transforms and their po
ssible use in removing badly non-Gaussian noise. Analytic and computational
results are presented to show that in the presence of highly non-Gaussian
noise, the coefficients of the nonlinear transform have much better propert
ies than traditional wavelet coefficients.