Three-dimensional object identification system using artificial neural networks

A new model-based three-dimensional object identification system using artificial neural networks is proposed in this dissertation. The proposed system consists of three parts: the geometric modeler, the aspect matcher and the aspect categorizer. The geometric modeler will build automatically three-dimensional identification model and two-dimensional input aspect description from a given CAD model and lower level image features. Topological information is also extracted and added to this model and description. Then, the aspect matcher will create and use the local compatibility table in guiding the process to find the matching faces in an input aspect and a three-dimensional object model. The matching is done on the modified Hopfield neural network. The aspect matcher also generates the invariant area patterns from matched faces. At the final stage, the aspect categorizer will accept the invariant area pattern and categorize the aspect viewing positions to store the aspect information of the identified object. To classify and store the aspect category, a modified ART-style neural network is applied. The important properties of the proposed identification system are automatic creation of identification model, invariant mapping of visible face area pattern and real time identification, learning and estimation of aspect categories.