Detecting salient curvature features using the local control of the feature support

Highly curved parts of object boundaries (curvature features) regularly correspond to characteristic image regions with high information content, which make them good candidates for object recognition and for establishing image correspondence in stereo or motion analysis. Unfortunately, computer vision procedures for detecting these features must deal with the conceptual problem of context dependency, since similar regions in different contexts may correspond sometimes to a salient curvature feature, and other times to noise or a deformed part of a straight boundary. In order to disambiguate such situations, a robust procedure should consider an appropriate neighbourhood for each candidate image location. In the proposed approach, the dimensionality of the search is reduced by preprocessing the input image by tuned edge detection and linking algorithms. Thus, curvature features can be searched only at points not ruled out in the preprocessing, while the complexity of the analysis is reduced from quadratic to linear. A representative subset of processing results is provided.