Linear Feature Extraction Based On Grouping Factors

Human vision has marvelous ability in extracting linear features from images, such as roads, rivers and so on. In this paper we present a new method to simulate this ability. Our method is based on some general grouping factors arising at two levels. At the first level, grouping factors are identified as direct bar-bar interaction and orientation interaction. Bar-bar interaction is shortranged and homogeneous. Orientation interaction is locally oriented and mediated by statistics of local visual context. At the second level, grouping factors are global geometric binding effects which arise from geometric redundancy reduction and thus are global effects. Based on them, we present an energy model. Then the extraction of linear features is generally formulated as combinatorial optimization. Since local, global interactions and local context effects all are included in it, the model may capture partially grouping ability of human vision systems. The experiments show that, without selecting any original points, our method can extract linear features from images robustly and quickly.