Super-resolution

Régis Destobbeleire

IMPA, Rio de Janeiro, Brasil

Training period responsible: Luiz Velho

07/04/2002-03/07/2002

Abstract

Sharpness is an important parameter in signal processing. Indeed, interpreting an image is determined by the possibility of extracting the information it contains. The sharper an image is, the more details it shows. If it is blurred, it will be hard to interpret it. Thus we are limited by the resolution of the image : if we want to zoom beyond this resolution, the necessary interpolation will generate a blurred image. In fact, it is impossible to create the missing information, which corresponds to the details we cannot see in the original image, and we would like to see in the zoomed one.

However, it is conceivable to guess this missing data. This would allow to increase the original resolution of an image : this is called super-resolution. I have chosen to study two different approches :

The first one, which was developed by William T. Freeman, consists in using as a reference a set of sharp images. This set allows us to learn the relation between high- and low-frequency bands in an image, and then to guess the missing high-frequency band of a blurred image.

The other one, developed by Aaron Hertzmann, uses as a reference a couple of images (the sharp and blurred versions of the same scene) to create, by analogy, the sharp version of another scene.

I implemented an algorithm described by W. T. Freeman in C++, as an interface, which permits to display the images and to control some parameters.

Some results are available here.
The report is available in pdf (3,51 Mb).