Object Boundary Segmentation Using Graph Cuts Based Active Contours

In this paper we propose an iterative graph cuts based active contours approach to segment an object boundary out of background. Given an initial boundary nearby the object, the graph cuts based active contour can iteratively deform to the object boundary even if there are large discontinuities and noise. In each iteration, the area of interest is a certain neighborhood of the previously estimated boundary. The problem is formulated as a multi-source multisink s− t minimum cut problem in that neighborhood and solved using node identification. The result of each step is globally optimal within the area of interest. On one hand, this area of interest enables the active contour to break away from the constraints of previously estimated boundary and isolated noise points in the image. On the other hand, our method generates multiple sources such that the result boundary should be long enough and hence overcomes the well-known shortcoming of the cost function of min-cut. It is shown that the proposed approach will converge to a final boundary. If this final boundary is not satisfactory, our approach is inherently suitable for interactive correction. As an advantage of using graph cuts, the proposed approach can be easily and seamlessly used to segment two dimensional(2D), three dimensional(3D) and possibly higher dimensional objects.