Most of my research falls under the umbrella
of
Machine Learning.
The focus of my
current research is on computer-based image segmentation,
which is a key step in computer vision.
Computer vision is concerned with the theory
for building artificial systems that obtain
information from images, i.e., machines that
can “see”. It is a relatively young field
that attempts to emulate the extremely
complex human visual system by using image
capturing equipment as the "eyes" and
computers and algorithms as the "brain". A
fundamental task of the human visual system
is the ability to find objects in an
image, i.e., segmenting an image into
semantically meaningful regions. Thus, over
the last four decades, the development of
image segmentation algorithms has been an
area of considerable research activity. Many image
segmentation algorithms have been
elaborated. However, strong segmentation,
which corresponds to a partitioning of an
image's pixels into regions that are
semantically meaningful to people, remains a
difficult and as yet largely unsolved
problem. Most current image segmentation
algorithms extract regions satisfying some
uniformity (homogeneity) criterion which is
based on low-level data-driven visual
features (e.g., color, texture). Those
algorithms perform well in narrow domains
(e.g., medical images, frontal views of
faces) where the variability of low-level
visual content is limited. Unfortunately, in
broader domains, homogeneous regions do not
necessarily (and usually do not) correspond
to semantically meaningful objects. This is
mainly caused by the disconnection between
low-level visual features and high-level
semantics, which is commonly referred to as
the semantic gap. Thus, the current
image-processing-based segmentation
paradigm, which relies exclusively on
low-level visual content, may have reached
its limits. To build strong image
segmentation algorithms that break through
the performance ceiling imposed by the
semantic gap, we must introduce high-level
semantic information into the segmentation
process. My current research deals with the
development of systematic image segmentation
strategies which, through the use of machine
learning, incorporate higher level knowledge
into the segmentation process.
My
Ph.D. dissertation investigated the
possibility of exploiting long-term learning
to improve the performance of content-based
image retrieval (CBIR). CBIR is an area of
intensive research. It aims at retrieval of
images from a database that are relevant to
a query image based on automatically derived
image features (e.g., color, texture). The
relevance of a database image to the query
image is proportional to the distance
between their corresponding points in the
feature space. Unfortunately, human notion
of similarity is usually based on high-level
abstractions such as activities, events, or
emotions displayed in an image. As a result,
images with high feature similarity to the
query image may be completely different from
it in terms of semantics. Thus, the semantic
gap is also an open challenging problem in
CBIR. Relevance feedback (RF) has been
proposed as a learning technique aimed at
reducing the semantic gap. It works by
gathering semantic information from user
interaction. Based on the user's feedback on
the retrieval results, the retrieval scheme
is adjusted. A disadvantage of traditional
RF approaches is that the captured knowledge
in one query session is not memorized and
the learning starts from ground up for each
new query. While short-term learning has
been widely used in the literature, less
research has been focused on exploiting
long-term learning.
The
focus of my master thesis was on the
development of a parallel genetic
algorithm
for the graph
theory/combinatorial problem of finding
Ramsey
Numbers.
Please look at my list of
publications to see what I am currently
thinking about.