A neural recognition architecture for composedobjectsGunther

We present an architecture for object recognition based on artiicial neural networks (ANN). The system can be trained on the holis-tic recognition of wooden toy pieces and aggregates composed of these pieces. However, the more complex aggregates become, the more diicult becomes holistic recognition. Therefore, after a \\rst glance" hypothesis by the holistic recognition module, the aggregate must be inspected visually for the single components. This can be done by a specialized holistic system, which is able to detect the basic toy pieces even within an aggregate. Another ANN, that can be looked upon as a model of the aggregate, can decide whether the geometric relations between the components found are correct. This approach is a step towards the integration of specialized holistic recognition modules to a recognition system for more complex aggregates.