Hierarchical merging of adjacent subtrees from Delaunay triangulation with centers of superpixels

Image segmentation is used in computer vision, medical imaging, and biological imaging to locate object boundaries and to group similar pixels together to form a set of coherent image regions. The important factors of clustering are similarity, proximity, and good continuation, which lead to visually meaningful segmentation. On the contrary, there are some problems of visual grouping such as over-segmentation, inaccuracy, and time-consuming tasks. Among the problems, we concentrate on reducing manual settings and avoiding the over-segmentation. In the paper, we propose a segmentation method which is merging hierarchically partial trees of superpixels. Given the superpixels, we determine the each center of superpixels to each node, and construct a Delaunay triangulation to compute which regions are adjacent. Similar regions are joined by using a similarity measure. An important chacteristic of the algorithm is its ability to reduce the over-segmentation and to preserve detail.