A multi-level invariant representation of digital images

Abstract

Recognizing a digital image from an object dictionary is a slow and memory consuming process. In this work the process was accelerated by using a compact geometrical model and a multi-level representation in conjunction with a database. An invariant representation of the image is used which is constant under object's position, orientation or size. The recognition system is based on a hierarchical level to accelerate the identification process; one level divides the image into a coarse and detailed model. The coarse one generates the overall feature descriptors used to minimize identification of unknown image through clustering, while the detailed model is used to precisely store the image such that an accurate template match is made possible. Another level of hierarchy is achieved by grouping objects of similar silhouettes into classes; once an unknown object is matched to a known class, it is further matched to a specific object within that class depending on some delineating physical dimension.