Study On Semi-supervised Spectral Clustering Algorithm And The Application In Image Segmentation

In the field of machine learning and data mining, getting amount of labeled data is expensive,but the unlabeled data are readily available. So the semi-supervised learning emerged, and itbecomes a hot research topic in the field of data mining because of it has the both advantages ofsupervised learning and semi-supervised learning.Spectral clustering algorithm is a kind of high performance computing approach, it cancluster in any shape of data space and converge to the global optimal solution. Adding thesupervised information to spectral clustering is semi-supervised spectral clustering. Thesupervised information can effectively improve the clustering effect, so it has important researchvalue. The main research of this paper is studying on semi-supervised spectral clusteringalgorithm and the application in image segmentation. First we propose a approach based onNSDR(Near Strangers or Distant Relatives) model, which is called Semi-supervise spectralclustering based on NSDR. Then we propose a constraint expansion approach, and propose aapproach that is called Semi-supervise spectral clustering based on constraint expansion. Finally,the new algorithms are used in image segmentation. The innovation of this paper lies in thefollowing aspects:(1)Propose a semi-supervised spectral clustering based on NSDR model. The algorithm usesthe structure assumption of data and supervised information proposing a similarity measuremethod to modify the similarity matrix, so the supervised information improve the result ofspectral clustering.(2)Propose a semi-supervised spectral clustering based on constraints expansion. Thealgorithm combines a approach of constraints expansion based on density with NSDR-SSC.Then the supervised information can improve the result of clustering more effective.(3)The two semi-supervised spectral clustering algorithms are successfully applied in theimage segmentation. It is challenging using spectral clustering and its related algorithms inimage segmentation. Because the spectral clustering algorithm requires to calculate the similaritymatrix and eigenvalues, so it has a high computation complexity in the large data set. In thispaper, we use the Nystr(o|¨)m approximation method to solve this problem.