Centerness Peak Based Clustering and Image Segmentation

Abstract

The density peak based clustering algorithm is presented by assuming that cluster centers are local density peaks, and utilizes local density relationship to detect cluster centers. This algorithm has been shown to be effective and efficient in some experiments. However, by studying the clustering mechanism in depth, we find that it may not be appropriate to treat density peaks as cluster centers in some cases. On one hand, the cluster centers obtained this way are often inconsistent with human intuition. On the other hand, local density difference across clusters is likely to influence the cluster center identification result. To relieve this problem, we present centerness as an alternative criterion of cluster center detection. The centerness criterion reflects to which degree the neighborhood of one data is filled with the nearest neighbors evenly, and is calculated with a histogram based method in our approach. By selecting cluster centers from centerness peaks, the clustering can be accomplished in a similar way as density peak algorithm. Our approach relieves the aforementioned problems of density peak algorithm, and performs well in experiments with synthetic and real datasets.