A COMPUTER VISION SYSTEM FOR GENERATING OBJECT DESCRIPTION

HSIEN-CHE LEE, Purdue University

Abstract

Computer vision is difficult both in problem formulation and in computational methodology. Traditional approaches either emphasize too much on the bottom-up process or on the use of complete a priori knowledge about the images under study. It is felt that a proper integration of the two approaches is necessary to design a general computer vision system. Based on this concept, a computer vision system is proposed and partially implemented. The emphasis is on the problem of dealing with imperfect segmentation and recovering the 3-D shape of an object from a single image. An image representation, called the Gray Level Geographic Structure image representation, is proposed for preliminary segmentation and preattentive visual search. Edge extraction, edge linking, and edge description procedures are implemented with efficient algorithms. Regions for describing surface orientations are then extracted from the resulted edge map. Special primitives, such as parallelograms, ellipses, symmetric arcs, and corners, are also extracted using different procedures. These primitives are treated as projections of some regular space contours. The relations between the space contours and their projections can be used to computer the range of possible surface orientations for an extracted region. Based on the computed surface orientations of each region, the 3-D shape of an object can be generated. Mathematically speaking, the information available in a single image is always ambiguous. Under certain regularity assumptions, local evidences about the shape of an object can often be derived. However, these local evidences, though rich in amount, are often inconsistent and not precise. A process called "Selective Confirmation" is proposed to resolve the ambiguities among the local evidences. The results of this process are a useful shape description of the object, and can be used as keys for associative recalls to retrieve proper object models and knowledge about them. the models and knowledge can then be used to guide the vision system to complete its interpretation processes.

Degree

Ph.D.

Subject Area

Electrical engineering

COinS