A Hierarchical Segmentation Algorithm Based on Mumford and Shah Model

This paper proposes a new hierarchical segmentation algorithm. Comparing with the previous work, the segmentation procedure is divided into two steps: a pre-segmenting stage and a tuning stage. By this way, some difficulties like the initialization of the level set and converging to local minima can be completely eliminated. To find an initial segmentation that will be applied as initial contours for tuning, a novel edge detection algorithm is introduced at the pre-segmenting stage, which is implemented based on the Mumford and Shah model in level set frame. Due to the contours obtained with the properties of level set functions, many active contour approaches may be used for tuning at second stage. However, aim at segmenting noisy images, only the piecewise smooth Mumford and Shah approach is taken into consideration in this paper. The resulting algorithm has been demonstrated by several cases.