A neural network system for two-dimensional feature recognition

Recognition of geometric features is important for automatic evaluation of part designs and development of process plans. This paper describes an implementation of a neural network for feature recognition in sheet metal parts created in a CAD system. One major part of the implementation is the development of a rotation-and translation-insensitive encoding scheme which extracts critical information from geometric features and candidate geometric loops. The successful implementation has led to a powerful system where end users can customize the domain of features that can be recognized. Training of the neural network memory is also achieved through a user-friendly graphic interface.