Clustering Methods Based On Normalized Cut And Watershed

Clustering analysis is a classical problem in machine learning. The inner relation between the objects can be showed by the results of the clustering methods based on the analysis of features of the objects. The goal of clustering analysis is to get various clusters. Objects belonging to the same cluster have the maximum similarity, and objects belonging to different cluster have the minimum similarity. By studying on clustering algorithms, we mainly propose three clustering methods, including spectral clustering based on similarity and dissimilarity, Hierarchical Clustering based on Normalized cut (HCN) and Three Dimensional Regional Watershed Algorithm (TDRWA).A spectral clustering based on similarity and dissimilarity method is proposed to further reflect the similarity between the samples in every clusters and the dissimilarity between different clusters. The spectral clustering based on similarity and dissimilarity method adopts a new clustering criterion function in which a distance matrix between samples is introduced into the spectral clustering based on the normalized cut. Experimants show that the proposed algorithm exhibits a high-performance.Classical distance-based hierarchical clustering methods can easily fall into a local optimum solution in non-convex datasets. In order to solve this problem, a new hierarchical merge clustering algorithm based on the normalized cut is proposed. In our algorithm, clusters which have the largest normalized cut should be merged firstly. Experimental results show that our algorithm is more efficient than classical ones on the non-convex distributed datasets.In practice, many exsiting techniques of medical image processing using clustering methods can not use the image information completely, and have large compuational complexity. In order to solve these problems, a TDRWA is presented for extracting the abdominal blood vessels in 3 dimensional CT images. TDRWA uses the 3 dimensional watershed and the region growing methods to process 3 dimensional CT images and extract abdominal blood vessels. The experimental results show that TDRWA could reduce the runtime and get a better result.