A framework for object representation and recognition

Abstract

The recognition task is transformed into simpler subtasks. Two assumptions are vital in this approach: (a) the object representation is pictorial, and (b) the parts of the object do not bear any information about the shape of the object. The aim is to find a framework which will make the problem of recognition easier. The recognition consists of two subtasks: classification of the object into its proper class and identification of the particular member of the class. The classification is performed on the basis of the object's iconic representation; the identification is based on the pattern representation. This fact is used to propose a multiresolution architecture which features classification of the whole object at only one resolution. It provides a framework in which the contemporary neural networks being applied to simple problems may be applied to real-world problems of visual object recognition. >