A self-organizing system for object analysis and identification

A parallel, self-similar network is proposed for the analysis and identification of possibly complex objects. The system allows one to define objects, or even conceptions at more abstract hierarchical levels, as equivalence classes. Based on this hierarchical, parallel, and distributed definition, the system constructs its networks. The system considers the object as an assembly of subparts or stiff components. Each stiff component can be defined by its geometrical contour and by surface characteristics (smoothness, texture, color, . . .). The entire object is defined hierarchically by the spatial and structural interrelations among the components. The system is able to recognize and analyze complex objects. Examples on real images are presented. It is also shown that the analysis can lead to an identification of specific instances of a complex equivalence class, as specific components and features of a new instance can be compared with several known instances. The identification of human faces is discussed as a test case