Automatic learning of structural models for workpiece recognition systems

A system for learning structural models for the recognition of partially occluded workpieces is described. The system is based on learning by showing, i.e., the models are constructed after some reference images of workpieces to be recognized have been presented to the system. Model learning is done by means of iterative optimization procedures: model description elements are selected, filter parameters are adapted to workpieces, and a strategy controlling the recognition procedure is determined. The system is implemented for learning 2-D models, but extension to 3-D model learning has been considered in the system design.<<ETX>>