Spectral Clustering Algorithm And Its Application In Image Recognition

Spectral clustering algorithm is a kind of new clustering algorithm.Compared with the traditional clustering algorithm, it can cluster on the arbitraryshape of sample space, has the advantage to converge to the global optima, andgradually become a hot research topic in the field of machine learning. Thespectral clustering algorithm mainly utilized the eigenvectors which come fromeigen decomposition of similarity matrix to cluster, so the influence analysis ofthe methods of constructing similarity matrix and how to construct a suitablesimilarity matrix for spectral clustering are of great significance to improve theperformance of spectral clustering algorithm.Firstly, in this paper the construction methods of similarity matrix for theexisting spectral clustering algorithms was summarized and collated, and therelated evaluation experiment for several common construction methods werecarried out to evaluate the influence on performance of spectral clusteringalgorithm with a variety of clustering performance evaluation metrics.Secondly, on the basis of similarity calculation model about sharedneighbor adaptive Gaussian kernel, combined with the concept of nodeimportance in social network analysis theory, a Gaussian kernel function wasconstructed based on the importance of shared neighbors, a kind of based onshared neighbor importance adaptive spectral clustering algorithm was proposed,and the experiments were carried out on some different datasets. The experimentresults show that in terms of clustering accuracy and stability and so on, theproposed algorithm obtains a better performance.Thirdly, with the application background about the recognition of onetraffic road health problems, the proposed algorithm was applied todigital image recognition system in this paper. The system mainly used the roadimage to recognize the cleaning condition of a crossroads over a period of time.Due to the image background is more complex and is difficult to extractedfeatures directly for target area. In this paper texture feature based onco-occurrence matrix (GCLM) were extracted with image pretreatment process such as the image segmentation, partitioning and so on, and then recognitionwith the proposed spectral clustering algorithm was preceded. The experimentresults show that the proposed algorithm has a better recognition ability andfault tolerance.Finally, the work of this paper and introduce some faced problems for thespectral clustering algorithm were summarized.