This paper addresses the automatic inference of a Gibbs distribution dedicated to segment grouping through relaxation labeling. The behavior of this method is studied through the detection of a road-like network from a noisy set of segments extracted from an image during a preprocessing step. Linking segments are added to this set to recover lost road parts. The whole segment set is organized in a relational graph and the road network restoration is modeled as a labeling process. The solution is defined as the labeling maximizing a Gibbs distribution constructed from a set of local costs computed for each graph clique. These cost functions, corresponding to interaction potentials, are learned automatically using multi-layer perceptrons. Supervised learning is performed over a training data set using only binary teaching output, “good” or “bad” configuration example. Several neural networks are used to overcome the problem of the variable complexity of clique configurations.
CITATION STYLE
Rivière, D., Mangin, J. F., Martinez, J. M., Chavand, F., & Frouin, V. (1998). Neural network based learning of local compatibilities for segment grouping. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 1451, pp. 349–358). Springer Verlag. https://doi.org/10.1007/bfb0033253
Mendeley helps you to discover research relevant for your work.