Multiscale image segmentation with a dynamic label tree

Automatic information extraction from satellite images is the base of remote sensing image archives with content-based query services. Pyramidal image models based on multiscale Markov random fields in combination with a texture model proved to yield good classification and segmentation results. The texture model is used for initial soft classification and then the optimal segmentation given the classification is found using a hierarchical process. Segment probabilities are calculated in a fine-to-rough analysis and segmentation is performed by a rough-to-fine decision algorithm. Previously proposed models optimise the strength of the dependencies in a fixed hierarchical structure. In the authors' model they allow the dependencies to switch, so that the hierarchical structure itself is optimised. Their model is exactly tractable, achieves very smooth segmentations, even at coarse scale, and can be fast computed.