The Study Of Muitiobjective Evolutionary Algorithm With Decomposition For Ciustering

Data mining is a simple process from which we can get hidden, unknown butuseful information and knowledge from a lot of incomplete, noisy, fuzzy and randomdata. As one of the mainstream techniques for data mining, clustering arouses more andmore attention. A large number of clustering algorithms have been proposed so far, butthey are designed only for special problems and users with imperfection theories andmethods. With the development of the internet, the data scale increasing. Butinsufficient of prior knowledge, makes that how to interpret those data, which are largescale, high dimension and unbalanced, become a difficult problem. The paper based onthe framework of multiobjective evolutionary, using optimized method to solveclustering problem. An optimization method named Multiobjective EvolutionaryClustering Algorithm with decomposition is proposed. The algorithm is successfullyused to artificial data clustering, UCI data clustering, texture image segmentation andSAR image segmentation. The main contribution can be listed as follows:1) Based on the framework of multiobjective evolutionary, a semi-supervisedmulti-objective clustering algorithm based on MOEA/D is proposed. Compared with thetraditional multi-objective evolutionary clustering algorithm, the proposed methoddecomposed a multi-objective clustering problem into several single-objectivesubproblems. Which makes the proposed algorithm has lower computational complexity.We have tested the algorithm in the special artificial data sets and UCI data sets.Compared with other five methods, the proposed method has a good performance.2) An improved multi-objective evolutionary clustering algorithm is applied toimage segmentation. An algorithm named multi-objective evolutionary clustering basedon decomposition for image segmentation is proposed. The method has been tested onsix texture images and three SAR images. Compared with other two methods, theproposed method has a good performance for image segmentation.3) A multi-objective evolutionary for synthetic aperture radar image segmentationwith non-local means denoising is proposed. In the processing stage, an improvednon-local means is ultilized to denoise the speckle noise of the SAR image. Then, oversegmenting object image into small regions by watershed segmentation algorithm. Themethod has been tested on six SAR images. Compared with other three methods, theproposed method has a good performance for SAR image segmentation. This research is supported by the National Natural Science Foundation ofChina(Grant Nos.61272279and61001202), the China Postdoctoral Science FoundationSpecial funded project (No.200801426), the China Postdoctoral Science Foundationfunded project (No.20080431228) and the Fundamental Research Funds for the CentralUniversities (No.JY10000902040).