Region Matching in Free-Form Surfaces Using Self Organizing Maps

This paper investigates the use of self-organizing maps (SOMs) to segment, describe and index free form surfaces. The overall objective of the research is to be able to take a surface region and search a database of surfaces to find matching regions. A free-form surface in this paper is described as a mesh of triangles and is not assumed to be easily parameterized. The three issues considered in this paper are the features to use, the size of the SOM and the different forms of training data used to train the SOM for indexing. There are numerous curvature based features that can be used to describe free-form surfaces and combinations of these are investigated to see if the matching performance is better than using just one feature type. The SOM is trained with the features to segment the free-form surfaces by clustering the triangular patches. The issue addressed is how the size of the SOM affects the clustering and hence segmentation. The indexing method proposed is dependant on the how the SOM is trained and three methods are compared. The three methods are: (1) training on the surface of interest, (2) training on the whole database of surfaces, and (3) training on a small set of parameterised surfaces. The results show that more than two surface features produce the best results, care must be taken when choosing the size of the SOM and training the SOM on a set of parameterised surfaces yields the best results. Results are shown for surfaces of faces that are essentially free-form and the goal is to be able to apply the techniques to searching databases of industrial parts.