Title page for ETD etd-04122004-155732

Type of Document

Thesis

Author

Sun, Donghu

Author's Email Address

donghsun@cs.fsu.edu

URN

etd-04122004-155732

Title

A Study Of Image Representations For Content-Based Image Retrieval

Degree

Master of Science

Department

Computer Science, Department of

Advisory Committee

Advisor Name	Title
Xiuwen Liu	Committee Chair
Anuj Srivastava	Committee Member
Daniel Schwartz	Committee Member

Keywords

spectral space
semantic analysis
intrinsic generalization sampling

Date of Defense

2004-04-05

Availability

unrestricted

Abstract

The performance of a content-based image retrieval system depends on the representation
of images. As a typical image consists of different objects, an image segmentation is needed
for more accurate representations of contents. The first part of this thesis describes a
generic image segmentation algorithm based on local spectral histograms of images. This
algorithm, demonstrated by experimental results, is shown to be effective for both texture
and non-texture images, and comparable to other segmentation algorithms. Due to the
time constraint of an image retrieval system, the second part of this thesis focuses on low
dimensional representations of images. By analyzing the semantics of commonly used linear
subspace representations through sampling their intrinsic generalizations, their limitations
are illustrated and a nonlinear representation, called Spectral Subspace Analysis (SSA) that
overcomes these limitations is proposed. In addition, to obtain optimal retrieval performance,
an algorithm for learning optimal representations is developed by formulating the problem as
an optimization one on a Grassmann manifold and exploiting the underlying geometry of the
manifold. Experimental results on different datasets show that both the SSA representation
and the learned optimal representations can improve retrieval performance significantly for
content-based image retrieval systems.

Files

Filename		Size	Approximate Download Time (Hours:Minutes:Seconds)
Filename		Size	28.8 Modem	56K Modem	ISDN (64 Kb)	ISDN (128 Kb)	Higher-speed Access
	donghu2004spring.pdf	1.12 Mb	00:05:11	00:02:40	00:02:20	00:01:10	00:00:05