Research On Multi-Objective Clustering Ensemble And Its Application

As an unsupervised classification technique,clustering is an important branch of data mining,which has been widely used in image segmentation and pattern recognition.So far,a lot of clustering algorithms have been proposed,which usually optimize single objective function.Recently,multi-objective evolutionary algorithm and ensemble learning technique have been employed for clustering,named multi-objective clustering algorithm and clustering ensemble.Compared with traditional clustering algorithm,these algorithms can get better clustering results.Based on these two clustering algorithms,researchers try to combine them and propose a new clustering algorithm,named multi-objective clustering ensemble,which performs better than these two algorithms.Against this background,a dual similarity based multi-objective clustering ensemble algorithm is proposed in this paper and is applied on image segmentation problem.The main work of this paper is summarized as follows:(1)A dual similarity based multi-objective clustering ensemble algorithm named DSMOC is proposed.In this algorithm,firstly,a clustering ensemble algorithm is modified and employed as crossover operator to generate new clustering partitions during optimizing process;secondly,a search process based on K-Means algorithm is introduced to generate new individuals.Experimental results show that DSMOC outperforms related algorithms and the two proposed improvements are both effective to improve the performance of DSMOC.(2)Apply the proposed method on image segmentation.For this purpose,a super-pixel segmentation algorithm is adopted to preprocess the image,which can improve the efficiency of the subsequent clustering segmentation algorithm.Then,DSMOC is employed to partition the super pixels.In order to reduce the complexity of this segmentation process,we simplify this method.Firstly,we cancel the random sampling process of DSMOC;secondly,we reduce the number of optimization iterations on the premise of ensuring the quality of segmentation.In the experimental part,we apply the algorithm on four natural images with real class information.The experimental results show that DSMOC can achieve better segmentation results than other clustering algorithms in terms of clustering index and visual segmentation effect.