Cluster-Oriented Image Segmentation Approach

A new approach for segmenting multi-dimensional imagery, named Cluster-Oriented Image Segmentation Approach (COIS), is developed in this dissertation. COIS approach includes two steps: (1) employing a clustering method, in COIS the peak-climbing algorithm is used to cluster the feature vectors with a small scale and (2) merging the clusters generated in the first step to make them correspond to the regions in the image domain. The second step is the core of the COIS approach. The operation in the first step is based on the pixel level and in the second step on the cluster level. COIS approach may suitable for the segmentation of images such as color images, texture images and remote-sensing images. The tests have shown the capability of the COIS approach in image segmentation.There are two main shortages in many traditional clustering based image segmentation approaches. One is these approaches attempt to get a satisfactory segmentation result based on just one clustering operation by finding a perfect clustering scale parameter. This means these approaches actually based on a hypothesis that after one clustering operation there is a linkage between clusters and image regions. It is not true in many cases because different image regions may need different scales. Another shortage is characteristics of clusters in image spaces usually do not be taken account. COIS approach overcomes these two shortages. The emphasis of the COIS approach is on merging the clusters generated in the first step, instead of the clustering itself because no linkage between clusters and image regions is made in the initial step. The characteristics of clusters in both feature space and image space are taken into account at the same time in the COIS approach. Only a rough scale is used in the first step and no scale used in the second step.During developing the COIS approach, this dissertation proposes: (1) Feature space based Closeness Index ( CI ) and Close Mate ( CM ). CI is used for measuring the relationship between two clusters, determining whether two clusters can be merged or not and CM provides a potential cluster into which another cluster maybe merge. (2) Image space based Gathering Index ( GI) which is used for measuring the capability of a cluster forming a region. (3) Image space based Capturing Index ( Cal) and Capturing Mate ( CaM). Cal measuring the relationship of two clusters, i.e., the degree of one cluster to becaptured by another one. Cal determines whether two clusters can be merged or not in image space. CaM provides a potential cluster into which another cluster maybe merge. These indices form the base of cluster merging in the second step of the COIS approach. The combinations of these indices in feature and image spaces generate different merging approaches that could be used to solve different issues in image segmentation. (4) An interactively merging method. (5) A method of segmenting by changing the cell size. (6) An evenness index (El) to measure the evenness of a cluster in the image space. It can be used to remove scattered points in a cluster. (7) A quick list-histogram building method. (8) Compression method of histograms. (9) Two new peak-climbing methods. (4) and (5) can be used independently or as aids in merging procedures.