A two-level probabilistic method for grouping edge-based descriptive primitives is proposed, At the lower level, a voting mechanism based on the Direct Hough Transform is employed in order to obtain a relational graph whose nodes correspond to a set of straight segments extracted from an edge image, The nodes are linked to each other according to geometrical relationships (i.e., parallelism, collinearity, convergence) among the detected segments, At the higher level, the grouping process consists of assigning a label to each node. Subsets of nodes with the same label are identified as consistent segment groups, A nonlinear cost function is used as a measure evaluating the goodness of every proposed configuration of labels, Such a measure is interpreted as an energy function related to the probability of a field configuration derived from the Gibbs distribution; this corresponds to modeling the label graph as a Markov Random Field (MRF), The energy function is formed by the interactions of local terms with precise geometrical significances, The label MRF is characterized by a multiple neighborhood system and, hence, by multiple cliques, A Simulated Annealing algorithm is used to find the best label configuration. The approach has been tested on a wide number of synthetic and real scenes and related results are provided to show the capabilities of the proposed approach. (C) 1996 Academic Press,Inc.
Grouping as a searching process for minimum-energy configurations of labelled random fields
FORESTI, Gian Luca
1996-01-01
Abstract
A two-level probabilistic method for grouping edge-based descriptive primitives is proposed, At the lower level, a voting mechanism based on the Direct Hough Transform is employed in order to obtain a relational graph whose nodes correspond to a set of straight segments extracted from an edge image, The nodes are linked to each other according to geometrical relationships (i.e., parallelism, collinearity, convergence) among the detected segments, At the higher level, the grouping process consists of assigning a label to each node. Subsets of nodes with the same label are identified as consistent segment groups, A nonlinear cost function is used as a measure evaluating the goodness of every proposed configuration of labels, Such a measure is interpreted as an energy function related to the probability of a field configuration derived from the Gibbs distribution; this corresponds to modeling the label graph as a Markov Random Field (MRF), The energy function is formed by the interactions of local terms with precise geometrical significances, The label MRF is characterized by a multiple neighborhood system and, hence, by multiple cliques, A Simulated Annealing algorithm is used to find the best label configuration. The approach has been tested on a wide number of synthetic and real scenes and related results are provided to show the capabilities of the proposed approach. (C) 1996 Academic Press,Inc.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
