Matching hierarchical structures for shape recognition

In this thesis we aim to develop a framework for clustering trees and rep- resenting and learning a generative model of graph structures from a set of training samples. The approach is applied to the problem of the recognition and classification of shape abstracted in terms of its morphological skeleton. We make five contributions. The first is an algorithm to approximate tree edit-distance using relaxation labeling. The second is the introduction of the tree union, a representation capable of representing the modes of structural variation present in a set of trees. The third is an information theoretic approach to learning a generative model of tree structures from a training set. While the skeletal abstraction of shape was chosen mainly as a exper- imental vehicle, we, nonetheless, make some contributions to the fields of skeleton extraction and its graph representation. In particular, our fourth contribution is the development of a skeletonization method that corrects curvature effects in the Hamilton-Jacobi framework, improving its localiza- tion and noise sensitivity. Finally, we propose a shape-measure capable of characterizing shapes abstracted in terms of their skeleton. This measure has a number of interesting properties. In particular, it varies smoothly as the shape is deformed and can be easily computed using the presented skeleton extraction algorithm. Each chapter presents an experimental analysis of the proposed approaches applied to shape recognition problems.