Neural network based learning of local compatibilities for segment grouping

Abstract

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.