Next:
References
Up:
Multi-Level Probabilistic Relaxation
Previous:
4 Application
We presented in this paper a general framework for applying the method
of probabilistic relaxation across different resolution levels: the
method allows the incorporation of information and measurements
pertaining to separate resolution levels, and proceeds from the coarsest
level to the finest level. As such, it allows the flow of information in
a more natural way than the classical probabilistic relaxation approach
where in successive iteration steps the information from coarser and
coarser scales is incorporated. We applied our formalism to the problem
of multiresolution segmentation of coloured texture images where a
dictionary of permissible region label configurations is introduced; the
method then proceeds in a bootstrap way by calculating the measurements
that characterise a certain class from the pixels that have been
assigned with high confidence to this class at the previous resolution
level. The method was demonstrated with the help of some composite
colour texture images where the measurements at each resolution level
are the Luv colours that make up each texture at that level.
The basic assumptions the theory relies on is the independence of
measurements concerning the individual pixels. This assumption is
reasonably correct when the measurements are the colour/grey values of
the pixels. It is clearly incorrect when the measurements refer to other
pixel attributes like for example texture. However, it is an important
convenience assumption, necessary to allow the development of the work.
Often in practice, probabilistic label assignments are iterated to
compensate for the error introduced by this assumption. For example, in
our case, the relaxation labelling could be iterated a few times at each
resolution level before the solution obtained was propagated to the next
finer resolution level. A stopping criterion then would be necessary,
which could for example be: the process stops when the number of pixels
changing their probabilities are a small fraction of total image pixels.
Next:
References
Up:
Multi-Level Probabilistic Relaxation
Previous:
4 Application
Dr. Majid Mirmehdi
Wed Jul 2 18:24:08 BST 1997