Image Segmentation with Hierarchical Mean Shift

Connected channels were found somewhere in the feature space in traditional mean shift method. Consequently, to accomplish the segmentation by only one trial often leads to unsatisfactory result, especially in weak edges or regions without Gaussian character. To solve this problem, the authors design a hierarchical segment method based on mean shift technique. Upon on the original sampling data, clustering is performed iteratively with different bandwidths on the centers gained previously. Thus, a tree structure is established between the nodes in different levels. According to the inheritance relationship, the leaf nodes are merged and classified into different categories finally. The method was implemented and tested in both gray and color images. Compared with the traditional mean shift method, the hierarchical method has advantage to reserve the detaisl in the same scale. Also, although the multiple time clustering is needed, the computation cost will not increase obviously due to the decreasing sample sizes and varied bandwidths. The hierarchical mean shift method has been proved to be promising in the experiments. But more theoretical analysis is required and the method also needs to be improved by more experiments.