HIERARCHICAL NEURAL NETWORKS FOR LEARNING 3-DIMENSIONAL OBJECTS FROM RANGE IMAGES

A free-form three-dimensional (3-D) object recognition system using ar tificial neural networks (ANNs) is described. The system is able to le am and recognize 3-D objects that have various surface shapes. The typ es of surface shapes the system is able to handle include not only pre defined surfaces such as simple piecewise quadric surfaces but also mo re complex free-form surfaces. The system utilizes ANNs to derive indu ced representations and inductive leaming of 3-D object classes. Start ing with range image processing, the surfaces of objects are segmented into surface parts by analyzing local shape features called surface/s phere intersection signatures (SSISs). Two layers of self-organizing f eature maps (SOFMs) are then used to learn those segmented surface par ts and their geometrical relationships. By finding corresponding neuro ns in the SOFMs for all pairs of surface parts appearing in the observ ed object, the object is described by a binary image that represents f iring states of neurons. The learning vector quantization (LVQ) networ k is used for learning and recognizing 3-D objects from objects' binar y image descriptions. The recognition performance of the system is dem onstrated using several objects. (C) 1998 SPIE and IS&T. [S1017-9909(9 8)00201-3].