Use of Clustering to Assist Recognition in Computer Vision

Abstract

In computer vision many problems are of non-deterministic polynomial time complexity. One of these problems is graph matching. Suboptimal solutions have been proposed to efficiently do graph matching. This thesis investigates the use of unsupervised learning to cluster structured graph data in polynomial time. Clustering was done on attributed graph nodes and attributed graph node-arc-node triplets, and meaningful results were demonstrated. Self-organizing maps and the minimum message length program Snob were used. These clustering results may help a suboptimal graph matcher arrive at an acceptable solution at an acceptable time. The thesis proposes some methods to do so, but implementation is future work.