Image Segmentation That Merges Together Boundary and Region Information

Abstract

Segmentation is a classical yet important problem in vision. Most of the previous works are either region-based or boundary-based. The two approaches own complementary merits – while the region-based one always produces closed boundaries, the boundary-based one involves primarily local operations and avoids the complexity of deciding how homogeneous a region and how inhomogeneous neighboring regions should be. In this paper, we propose a new solution mechanism that makes use of both cues. We use the boundary processing and a particular field model to come up with a number of coarse, initial closed boundaries about the image first. Such coarse boundaries will then, through an adaptation of the Four Color Theorem, serve as the initialization to a level-set method-based minimization that acts on the intensity distribution of the image, and allows the final crispy segmentation result to emerge. Compared with the existing solutions, our method requires no initialization from the user, and the automatically extracted closed contours do provide guidance to derive more optimal and smoother segmentation result. Experimental results with some benchmarking image-sets show that the proposed solution could deliver accurate segmentation boundary.