Date of Award

12-2007

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Electrical Engineering

Advisor

Hoover, Adam

Committee Member

Muth , Eric

Committee Member

Walker , Ian

Committee Member

Birchfield , Stan

Abstract

In this dissertation a novel bottom-up computer vision approach is proposed. This approach is based upon quantifying the stability of the number of regions or count in a multi-dimensional parameter scale-space. The stability analysis comes from the properties of flat areas in the region count space generated through bottom-up algorithms of thresholding and region growing, hysteresis thresholding, variance-based region growing. The parameters used can be threshold, region growth, intensity statistics and other low-level parameters. The advantages and disadvantages of top-down, bottom-up and hybrid computational models are discussed. The approaches of scale-space, perceptual organization and clustering methods in computer vision are also analyzed, and the difference between our approach and these approaches is clarified. An overview of our stable count idea and implementation of three algorithms derived from this idea are presented. The algorithms are applied to real-world images as well as simulated signals. We have developed three experiments based upon our framework of stable region count. The experiments are using flower detector, peak detector and retinal image lesion detector respectively to process images and signals. The results from these experiments all suggest that our computer vision framework can solve different image and signal problems and provide satisfactory solutions. In the end future research directions and improvements are proposed.

Share

COinS