A new approach to machine-based perception of monocular images

A novel approach to machine-based visual perception of monocular images is described and initial recognition results are presented. The approach uses a recursive procedure to generate a series of reconstructed versions of the raw video image. The procedure is motivated by certain perceptual organization functions of the human visual system. Recognition of object categories is attempted at each step by comparing the newly generated regions to stored object categories. Category retrieval is carried out using a software-based content-addressing scheme which provides access to complete object representations in memory using incomplete portions of the representation. The category-representation scheme is sufficiently general to allow a variety of poorly correlated images of specific category examples to be represented and recognized by a single general category representation. These properties are illustrated using a group of distorted, defective, and idealized images of ASCII A´s and handguns