Fall Seminar Series - October 20, 2005
University of Minnesota
School of Statistics
College of Liberal Arts
Apex
Blind Deconvolution of Hubble Telescope Imagery and the Use of Levy
Stable Laws
Alfred S. Carasso
National Institute of Standards and Technology
Thursday, October 20, 2005
3:30 PM, 115
Ford Hall
Minneapolis, East Bank Campus
Social at 3:00 PM, 300 Ford Hall
Abstract
Blind deconvolution seeks to deblur an image without knowing
the cause of the blur. This problem is fraught with mathematical
difficulties, as severe ill-conditioning is compounded with
non-uniqueness of
solutions. Few reliable algorithms are known, and most
approaches involve time-consuming iterative procedures.
Recently, a new non-iterative approach to blind deconvolution was
formulated, based on suitably restricting the allowable class of blurs,
and
using Fast Fourier Transform techniques. This so-called APEX
method can process high resolution 1024 x 1024 imagery in quasi
real-time, a monumental breakthrough for the class
of images for which the method is applicable. However, not all images
can be usefully enhanced with the APEX method.
In constructing the APEX method, selection of the allowable
class of blurs was guided primarily by considerations of pure
mathematics. A class of point spread functions was sought,
that generalized the Gaussian density, possessed appropriate
semigroup properties, and a relatively
simple Fourier space representation. These properties
result in a tractable deconvolution problem
that can be solved in slow motion. The class of radially
symmetric L\'{e}vy stable characteristic functions
\hat{h}(\xi,\eta)= exp\{-\alpha (\xi2+\eta2)^{\beta}\},
\qquad \alpha > 0,~~0 < \beta \leq 1,
is the simplest example of a broad class of blurs with the requisite
mathematical properties. That class is currently the basis for the APEX
method.
The APEX method has been found surprisingly effective in quite diverse
applications, including MRI and PET brain scans, scanning electron
microscopy, and most recently, Hubble space telescope color
imagery. In most cases, the detected point spread functions that
successfully enhance these images have very low exponent $\beta$, and
are very
far from Gaussian. The reasons behind these successful applications
remain an open problem.