IMAGE SEGMENTATION WITH COMPLEMENTARY USE OF EDGE AND REGION INFORMATION

Abstract

Two familiar approaches to image segmentation are the salient contour extraction approach and the closed-contour deformation approach. The former uses Gestalt laws to link individual edge elements and construct segmentation boundaries. However, it is often difficult to have both closure and precision of the boundary addressed at the same time. The latter starts with a closed contour and deforms the contour to localize the segmentation boundary more precisely whilst maintaining the closure. The approach does not have the closure problem, but how to assign a proper initial contour for it remains an open issue. In this work, we propose a scheme that puts together the two approaches to let them work complementarily. Specifically, we design a salient contour extraction process that extracts a proper initialization of the closed contours; the process looks into edge evidence and proximity to the desired segmentation boundaries. Then, a region-based active contour in a level set formulation is adopted to refine the contour position to locate the segmentation boundaries more precisely. The scheme requires neither manual input on contour initialization nor prior knowledge about the imaged scene. Experiments on extensive benchmarking image-sets are presented to illustrate the performance of the scheme.