Comparison with a stochastic model
We have compared the results we obtain from the proposed variational model with
the ones obtained using a stochastic based on Markov Random Fields
theory (cf. [2] for instance).
Let be the solution of classification, i.e. an image of labels, and let represent the observed data. We are
looking for the estimation of
by
maximizing (Maximum A Posteriori) the probability P(/).
Using Bayes rule, and according to the properties of Markov Random
Fields, the solution is found thanks to :
The Energy E contains a Gaussian data term, and a
regularizing term (A Priori) known as a Potts
regularization term. N is the number of pixels, and are respectively the label assigned to the
site s and the observation at site s. As for the
proposed variational model, the label value assigned to
the site s is one of the values ,
i=1..M. The set C is the one of considered cliques. Parameter
represents the weight of the regularization term and
the operator is the Dirac distribution defined as
stands for the temperature parameter.
The minimization of E is operated through a simulated annealing (with a Metropolis dynamic).
- Introduction
- Proposed model
- Algorithm
- Results : synthetic image
- Results : first satellite image
- Results : second satellite image
- Conclusion
- Bibliography