Data Clustering and Graph-Based Image Matching Methods

This thesis describes our novel methods for data clustering, graph characterizing and image matching.
In Chapter 3, our main contribution is the M1NN agglomerative clustering method with a new parallel merging algorithm. A cluster characterizing quantity is derived from the path-based dissimilarity measure.
In Chapter 4, our main contribution is the modified log likelihood model for quantitative clustering analysis. The energy of a graph is adopted to define the description length to measure the complexity of a clustering.
In Chapter 5, our main contribution is an image matching method based on Delaunay graph characterization and node selection. A normalized Euclidean distance on Delaunay graphs is found useful to estimate pairwise distances.