A neural network for shape recognition

A neural network that has the capability for viewer-independent recognition of occluded, complex three-dimensional objects is introduced. The technique is based on a set of object-dependent points known as critical points. These points are derived from a structure known as the concavity tree, which is a unique representation for planar shapes. Shapes or objects are compared and identified based on feature vectors formed from the critical point sets. Each feature vector is composed of exactly two critical points where the subsequent feature vectors are computed in succession along the contour of the shape. Finally, the feature vector representation is a ratio expression utilizing two successive feature vectors.<>