Super Resolution is an artificial increasing
of the actual image resolution. It attempts to reconstruct
the original scene that gave rise to the image.
The most frequently used ways to solve this
problem are different kinds of image interpolation. Bicubic
interpolation is probably the most popular one. The interpolation
procedures are presented in many image processing software
Interpolation is not a good solution. The
problem is that interpolation can not ensure the effective
restoration of the highest frequency part of the image spectrum.
This means that many of the smallest details and boundaries
of a complicated configuration can not be restored by using
interpolation. The image obtained by using interpolation
as a rule will not be sharp, and often it may be smoothed.
This is a significant disadvantage of interpolation.
Super-resolution means not only a formal
enlargement of the image. First of all it means extraction
of the unknown part of image spectra in the highest frequency
domain. Only such a solution ensures preservation of image
details and extraction of the smallest details that are invisible
on the input image.
Our super-resolution algorithm is based on
this approach. It extrapolates the image spectra to the highest
frequency domain. Comparison of the images resampled by the
interpolation and our super- resolution show that the latter
is much more powerful both from the subjective point of view
(visual impression) and the objective one. The latter means
that the high frequency part of the spectrum corresponding
to the interpolated image is very poor, while one corresponding
to the image obtained by our super-resolution contains a
lot of useful information.
The solution of the super resolution problem is based on
the extrapolation of the orthogonal image spectra (Fourier,
Cosine or Walsh) to the unknown highest frequency domain.
This approach is implemented by using iterative approximation
of the unknown high frequency part of the orthogonal spectra,
and final correction of the obtained image with increased
resolution, using multi-valued filters.
|Comparison chart of spectra of the
supersampled image to the images that are the results
of different interpolation techniques
(red -original, blue - our supersampling,
green - interpolation)
(click on to see supersampled image,
make sure you don't have pop-ups disabled)