KL Divergence-Based Fuzzy Cluster Ensemble for Image Segmentation.

Huiqin Wei,Long Chen,Li Guo

doi:10.3390/e20040273

Abstract

Ensemble clustering combines different basic partitions of a dataset into a more stable and robust one. Thus, cluster ensemble plays a significant role in applications like image segmentation. However, existing ensemble methods have a few demerits, including the lack of diversity of basic partitions and the low accuracy caused by data noise. In this paper, to get over these difficulties, we propose an efficient fuzzy cluster ensemble method based on Kullback–Leibler divergence or simply, the KL divergence. The data are first classified with distinct fuzzy clustering methods. Then, the soft clustering results are aggregated by a fuzzy KL divergence-based objective function. Moreover, for image segmentation problems, we utilize the local spatial information in the cluster ensemble algorithm to suppress the effect of noise. Experiment results reveal that the proposed methods outperform many other methods in synthetic and real image-segmentation problems.

Highlights

Image segmentation has become increasingly important in a wide variety of applications like biomedical image analysis [1,2,3,4] and intelligent robotics [5]
We first propose an efficient fuzzy cluster ensemble method based on KL divergence ( Fuzzy Cluster Ensemble Based on KL Divergence (FCE_KL) )
We illustrate the problem of combining multiple clustering operations and propose an efficient fuzzy cluster ensemble method based on KL divergence

Summary

Introduction

Image segmentation has become increasingly important in a wide variety of applications like biomedical image analysis [1,2,3,4] and intelligent robotics [5]. There is a growing number of methods available for image-segmentation problems over recent years [1,2,3,4,5,6,7,8]. In contrast to hard segmentation methods, the fuzzy ones could retain much more information from the original data [9,10,11]. The fuzzy c-means (FCM) clustering algorithm is the best known one in fuzzy segmentation methods [12]. The standard FCM does not consider any spatial information in the image context, and suffers from high sensitivity to noise. With so many available algorithms, one may obtain very different clustering results for a given dataset. Most of existing algorithms require the specification of some parameters to obtain a decent grouping of the data

Results

Discussion

Conclusion