The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence

"GRAPH BASED IMAGE SEGMENTATION"

By

Mr. Jingdong WANG


Abstract

Image segmentation refers to a process of dividing the image into disjoint
regions that were meaningful. This process is fundamental in computer
vision in that many applications, such as image retrieval, visual summary,
image based modeling, and so on, can essentially benefit from it. This
process is also challenging because the segmentation is usually subjective
and the computation is highly costly.

Recent researches have witnessed great successes of graph based image
segmentation. In graph based approaches, one of the key issues is graph
prior in statistics terms. In this thesis, we address this problem for
image segmentation from the following perspectives:

1) We investigate the prior design in the conventional pairwise graph
approaches. We design a joint graph prior in the joint segmentation for 3D
clustering and 2D image segmentation. This problem makes intensive use of
the current states of the art in both unbiased and biased graph
partitioning.

2) We generalize the pairwise graph to the hypergraph that can model
multiple-wise relation between the data points. We propose a biased
hypergraph partitioning approach, called Linear Neighborhood Propagation,
which leads to a linear system with an efficient solution.

3) We degrade the cyclic graph to a connected acyclic graph, i.e. tree. We
propose a tree partitioning method based on the normalized cut criterion,
called Normalized Tree Partitioning. It runs in linear time, and
potentially results in superior performance over spectral graph
partitioning due to less approximation.


Date:			Thursday, 23 August 2007

Time:			10:00a.m.-12:00noon

Venue:			Room 3416
			Lifts 17-18

Chairman:		Prof. Robert Ko (BICH)

Committee Members:	Prof. Long Quan (Supervisor)
			Prof. Helen Shen (Supervisor)
			Prof. Chiew-Lan Tai
			Prof. Dit-Yan Yeung
			Prof. Ajay Joneja (IELM)
			Prof. Michael Lyu (CSE, CUHK)


**** ALL are Welcome ****