Clustering-based Image Segmentation And Image Registration

The task of image segmentation is to divide an image into a number ofnon-overlapping regions, the pixels in which have same or similar characteristics. Theclustering-based image segmentation is a popular method in the field of segmentation.Image segmentation has great influence on all processes that refer to extracting andanalyzing objects, especially in military and medicine affairs. The purpose of imageregistration is to correct the geometric transformation of two or more images taken bydifferent sensors, or at different times, or from different viewpoints. Imagesegmentation and image registration are always hot topics in the field of imageprocessing. Clustering can be used as an effective segmentation method in field ofimage segmentation and image registration. Therefore, clustering method,clustering-based image segmentation and registration are investigated in this thesis.Details are described as below:Firstly, a novel two-phases clustering algorithm with density exploring distancemeasure was proposed. In the first phase, the proposed algorithm applied the fast globalk-means clustering algorithm. The prototypes of all clusters, representatives of pointsbelonging to these clusters, were regarded as the input data set of the second phase.Then all the prototypes were clustered according to a density exploring distancemeasure which makes data points locating in the same structure possess high similarityfor each other.Secondly, an image segmentation method based on the two-phases clusteringalgorithm and watershed segmentation was proposed. It was a combination oftwo-phases clustering and watershed segmentation. Marker driven watershed was usedto segment original image, and the two-phases clustering method was regarded assecondary segmentation. The three sub-band energy of discrete wavelet transform andgray level were extracted as feature vector of the image.Finally, an automatic image registration method, which combines imagesegmentation through fuzzy clustering as spatial restraint with SIFT was proposed.Seven moment invariants were extracted as features of objects obtained by imagesegmentation via the fast generalized fuzzy c-means algorithm. Then, we detected theSIFT keypoints in corresponding matching regions and combined them with thekeypoints obtained from the whole image, which was followed by a robust outlierremoval procedure and figured out the transform parameters.