James G. Nagy (Emory University), Veronica Mejia-Bustamante (Emory University)
Keywords: Imaging
Abstract:
Obtaining high resolution images of space objects from ground based telescopes involves using a combination of sophisticated hardware and computational post processing techniques. An important, and often highly effective, computational post processing tool used at AMOS is multi-frame blind deconvolution (MFBD) for image restoration.
One difficulty with using MFBD algorithms is that the nonlinear inverse problem they are designed to solve may have many local minima. Standard optimization methods that use the gradient to search for a minimum (e.g., the conjugate gradient method) may get trapped in a local minimum, resulting in a less than optimal restored image. One approach to get around this difficulty is to run the algorithm several times with different initial guesses, which then results (hopefully) in computing several different local minima. A pseudo global minimum is then found by determining the best of these local minima. There are several disadvantages to this approach, including the extensive cost for large scale problems.
In this paper we consider an alternative, computationally less expensive approach, based on preconditioning. Preconditioning is used widely in science and engineering to accelerate convergence of iterative optimization methods, but it is usually applied to convex problems that have only one local minimum. However, recently it has been observed that for a nonlinear inverse problem arising in inverse scattering that preconditioning can have the dual advantage of improving the rate of convergence and reducing the problem of becoming trapped in local minima. We investigate adapting this approach to MFBD. The most effective approaches are designed using information specific to the application, so we consider several schemes appropriate for use in MFBD algorithms, and for imaging of space objects.
Date of Conference: September 1-4. 2009
Track: Imaging