Segmentation and grouping of object boundaries using energy minimization

Traditionally, vision systems have largely relied upon the object boundaries extracted from images to accomplish recognition. Most contour detection algorithms, however, suffer from the fact that the extracted object boundaries are usually broken and incomplete due to poor imaging conditions and/or occlusions. In this paper, we describe an algorithm to perform curvilinear grouping of image edge elements for detecting object boundaries. The method works by generating hypotheses and selecting the best one. A neighborhood definition based on Delaunay graph is used to keep the number of hypotheses generated small. An energy minimizing curve is fit to the generated hypotheses to evaluate the grouping and locate discontinuities.