Edge Contours Using Multiple Scales

Abstract

An edge linking agorithm is presented that first computes gradient magnitude then uses non-maxima suppression to reduce the search space. The gradient magnitude and direction information is used to assign weights to paths through the set of points, and the best path is chosen. The single scale algorithm uses a depth first search, but each point is allowed to occur in only one subtree. This algorithm is extended to one which links edge points detected at multiple scales. A second variation is presented which uses the gradient information at multiple scales in the non-maxima suppression operation. A modified version of the single scale algorithm then links these edge points into contours. The first multi-scale method fills in gaps in single-scale contours, however the second method provides better localization and separation of nearby edges. Both improve detection of edges over the single-scale algorithm without introducing the noisy edges detected at the smaller scales.