Mats Matsson, presents his master thesis

Image Super-Resolution using Gaussian Markov Random Fields

Abstract

There are many scenarios where greater detail in an image is desired
than what is possible with the available equipment, e.g. in video
surveillance. A super-resolution image is an image with greater
resolution estimated from images with lower resolution. In this thesis
we examine the use of Intrinsic Gaussian Markov Random Fields (IGMRF)
to construct a super-resolution image from a number of low- resolution
images.
	
We describe an approach to this based on the EM-Algorithm with
stationary registration noise for the low-resolution images and a
stationary IGMRF prior distribution for the super-resolution image,
and then try to extend this to a non-stationary model for both. A non-
stationary image prior has been tried by others. The non-stationary
registration noise is something we believe is a novel approach.
	
The methods have been evaluated with simulated images, with good
results, and real images, with no improvements in detail. We found
that a non-stationary registration noise approach work, but had
unpredictable results using a non-stationary prior distribution.
	
The lack of improvement in detail for the real images may be due to
that we model the camera as a pin-hole camera, i.e. assume the lens is
perfect. Our limited modeling of color images is also something that
may improve future results.