Associative memory for geon-based object identification

Online inspection for defects in manufactured parts is a challenging problem and a crucial issue for any production process. Precise and extended CAD data describe the manufactured part under inspection and can be stored in a database in order to perform comparisons between the measured observed part (with a 3D laser camera, for instance) and its CAD description. The problem is very slow access time to find the right CAD model, because of the huge amount of data necessary to represent even a very simple manufactured piece. We propose a fast access to the database based on a vision-derived representation of the part, much less precise than the CAD representation but which offers a very low storage requirement. CAD representations of the models are converted off-line into a vision-based representation used to train a neural network. Volumetric projected components (geons) and their interrelationship are extracted from the digitized image of the piece under inspection and input to a neural network whose outputs point to the most probable models in the large database. Corresponding CAD data is then available for the inspection module