Theoretical Statistics and Mathematics Unit, ISI Delhi

July 17, 2015 (Friday) ,
3:30 PM at Webinar

Speaker:
Kaustav Nandy,
Indian Statistical Institute, Delhi

Title:
Deconvolution using natural image priors

Abstract of Talk

Blurring of images is a common phenomenon, typical examples being
astronomical images that may be degraded due to atmospheric factors or
telescope optics, and photographs that may be blurred due to motion of
the subject or camera shake during relatively long exposures.
Recovering the underlying image from an observed blurred image is an
interesting inference problem. As the blurring process is modeled as a
convolution of the underlying image and a `blur kernel' or `point
spread function', this problem is usually referred to as the
deconvolution problem. The problem is easier to solve when the blur
kernel is known, as is typical in astronomical images, and reasonable
solutions have been available for several decades. A more difficult
version of the problem is blind deconvolution, where the blur kernel
is unknown, as in photographs of natural scenes. Considerable progress
has been made recently by using priors on the space of natural images.
In this talk, we give an overview of the deconvolution problem and
present some preliminary work in trying to understand the blind
deconvolution problem and hopefully obtain better solutions than those
currently available.