A Graph-Based Representation to Detect Linear Features

摘要

Graph-based representations of the scene are well adapted to introduce high-level knowledge in image segmentation. The problem consists then in searching the graph configuration or labeling minimizing some cost function. In the case of local relationships between the graph nodes, the Markovian frame- work and simulated annealing algorithms provide some answers to this question. We are interested in this paper in the automatic or semi-automatic detection of linear structures like Roads or hydrological networks in satellite radar images. Using a graph of segments and introducing Local contextual properties of these networks, a Markov Random Field is defined to perform the Detection. Interaction choice relies on a priori knowledge on the usual aspect of the linear objects to Detect. Results are presented for real radar images both for road and hydrological networks.