A Graph Labelling Approach for Connected Feature Selection

摘要

Many authors have already proposed linear feature extraction algorithms. In most cases, these algorithms can not guarantee the extraction of adjacency relations between extracted features. Object contours appearing in the analyzed images are often fragmented into non-connected features. Nevertheless, the use of some topological information enables to reduce substantially the complexity of matching and registration algorithms. Here, we formulate the problem of linear feature extraction as an optimal labelling problem of a topological map obtained from low level operations. The originality of our approach is the maintaining of this data structure during the extraction process and the formulation of the problem of feature extraction as a global optimization problem.