May. 4 2005 2:00PM

May. 4 2005 3:30PM

souhrn:
Most remote-sensing images exhibit a clear hierarchical structure which can be taken into account by defining a suitable model for the unknown segmentation map. To this end one can resort to a tree-structured MRF model (TS-MRF), which describes a K-ary field by means of a sequence of M-ary nested MRFs, each one corresponding to a node in the tree, where M is variable from node to node. Each of such local fields, which is supported by an irregular domain shaped form the ancestor fields, can be designed independently form each other by defining own parameters whose estimate is locally achievable.
Dealing with supervised segmentation problems, where the tree structure of the model which better fits with the data is itself known a priori, the TS-MRF based algorithms provide quite promising results w.r.t. conventional MRF-based classifiers. On the other hand, in an unsupervised framework, thanks to the recursive formulation of the proposed models, they can be progressively built, split-by-split, by adding new local field components at each step if needed. In such a way the cluster validation problem can be addressed in a straightforward manner by controlling the tree growth with test parameters local to each node which indicates whether it needs to be split or not.