The Research And Application Of Spectral Clustering Algorithm Based On Neighbor Propagation And Density Information

Spectral clustering algorithm is a recently developed clustering algorithm based on graphtheory, compared with traditional clustering algorithms, has obvious advantages. Spectralclustering algorithm is not only simple, but also to identify the arbitrary shape of the samplespace, and can converge to the global optimal solution, is very suitable for practical problems,has become a research focus in the field of machine learning and pattern recognition. Sincethe affinity matrix of spectral clustering has a direct impact on the clustering performance.Therefore, this paper research on this issue, and has proposed three affinity matrixconstruction methods and corresponding spectral clustering algorithms. Then, we will applythe improved spectral clustering algorithm for image segmentation. Experimental results showthe effectiveness and superiority of our algorithm. Specifically, the main work is as follows:(1) We propose a second updating sample similarity between pairwise subsets metholdbesed on neighbor propagation and mode merging technology. First, we will updata thesimilarity between samples from different subsets to be the maximum similarity betweensamples from the two subsets, sothat the degree of closeness among smples are enlarge orcompress relatively; then we will merge the subsets whose number beyond the number ofclustering, and further updata the affinity matrix. Finally, using the new affinity, we will getan improved spectral clustering algorithm (NPMM-SC). The experimental results show thatthe secondary updated affinity matrix is more advantageous to obtain the correct clusteringresults; NPMM-SC algorithm is feasible and effective.(2) A kind of local density estimation and neighbor propagation based spectral clusteringalgorithm (LDENP-SC) was proposed. In this algorithm, the local density of the samples isfirst estimated and regarded as a new dimension of the dataset. Then, the similarity matrix isupdated by using neighbor propagation and is used to spectral cluster. Also, a simple localdensity estimation methodology was proposed in this paper. Moreover, based on propagationalgorithm, a new method for updating the similarity of the samples in different subsets wasadopted so as to update and get more actual similarity matrix. The experimental results showthat LDENP-SC can almost obtain the ideal similarity matrix and accurate clustering results,and it has good generalization ability and robustness for a certain range of parameters.(3) Based on the local density estimation and union feature, we propose a new method ofSAR image segmetation. The method first uses density estimation to calculate the pixelsdensity, and draw the corresponding density image and so on; we replace some area ofenhanced anti-density image and get the image for coarse segmentation; after the coarsesegmentation, we can get multiple regions which separated each other, and then we regardthese regions as sample points also feature extraction. Fanally, we can constructe thesimilarity matrix of sample points for spectral clustering, and achieve an algorithm(LDEUF-SC). The experimental results show that LDEUF-SC can effectively segment theSAR images and aerial images, the regional error less.