According to our Special Topics analysis of face
recognition technology over the past decade, the paper
"Face recognition using Laplacianfaces" (XF He, et
al., IEEE Trans. Patt. Anal. Mach. Int. 27[3]: 328-40,
March 2005) ranks at #14 on the list of most-cited papers
over the past decade. It is also a core paper in the
Research Front Map on
Face Recognition, and a Highly Cited Paper in
the field of Engineering in
Essential Science IndicatorsSMfrom
Thomson
Reuters.
This month, ScienceWatch.com
talks with lead author Professor Xiaofei He, who hails
from the College of Computer Science at Zhejiang University
in China, about this paper and its impact on the
field.
Would you please describe the significance of
your paper and why it is highly cited?
Our paper is the first to introduce manifold learning techniques to the
face recognition field. We were exploring the geometrical and topological
structures through a graph model. We found a set of basis functions, called
Laplacianfaces, which are linear approximations to the eigenfunctions of
the Laplace-Beltrami operator on the face manifold. The Laplacianface
approach allows more accurate estimation of face variations, such as
expression, illumination, and pose.
In addition, our paper systematically discusses the connection between our
approach and canonical approaches, such as Eigenface and Fisherface, and
shows that these approaches can be unified into a graph-embedding
framework. The presented framework of analysis leads to further research
and discussions. Our work provides experimental evidence that the intrinsic
manifold structure can significantly improve the recognition rate.
How did you become involved in this research, and
were there anyparticular successes or obstacles that
stand out?
I did my B.S. in computer science at Zhejiang University, China, and my
Ph.D. in computer science at the University of Chicago.
I have been working on manifold learning and face recognition since my
Ph.D. study. Since 2000, there has been a lot of interest in geometrically
and topologically motivated approaches to data analysis. After noticing the
close relationship between manifold learning methods and canonical
dimensionality reduction methods (such as Principal Component Analysis and
Linear Discriminant Analysis) which have been applied to face recognition,
I immediately started to think about how to use manifold structure to
improve face recognition performance. This eventually led to this highly
cited publication.
Where do you see your research and the broader
field leading in the future?
In November 2007, I moved from Yahoo! Research Labs in Burbank, where I was
a research scientist, to Zhejiang University in China, where I have
accepted a position as a full professor at the College of Computer Science.
I am currently leading a research group and continuing to work in the field
of face recognition and manifold learning.
"The Laplacianface approach allows
more accurate estimation of face variations,
such as expression, illumination, and
pose."
It is now a general belief that the images of human faces lie on
low-dimensional manifolds. The broader direction is thus discovering more
complex geometrical and topological properties of the face manifold, such
as curvature and homology group, from limited random samples. Furthermore,
it would be interesting to study how to use differential operators to
explore the meaningful structure of the functional space on the face
manifold from which we are looking for the optimal classifier.
These efforts would improve our understanding of the face recognition
problem from a geometrical perspective and may lead to a more accurate and
robust recognition algorithm, which is invariant to pose, illumination, and
expression.
What are the implications of your work for this
field?
The results of our research have significant implications for face
recognition, pattern classification, and manifold learning. With regard to
face recognition, we showed that the local manifold structure could better
describe the face variations. We also presented a graph embedding framework
and a novel linear manifold learning algorithm—Locality Preserving
Projections (LPP), which impacts our view of canonical dimensionality
reduction algorithms, such as Principal Component Analysis and Linear
Discriminant Analysis. LPP is a general tool for exploratory data analysis
and has been applied by other researchers to a wide range of problems such
as human age estimation, human action recognition, and graph pattern
matching.
Xiaofei He, Ph.D.
College of Computer Science
Zhejiang University
Hangzhou, People's Republic of China
Two-dimensional linear embedding of face images by Laplacianfaces. As can
be seen, the face images are divided into two parts, the faces with open
mouth and the faces with closed mouth. Moreover, it can be clearly seen
that the pose and expression of human faces change continuously and
smoothly, from top to bottom, from left to right. The bottom images
correspond to points along the right path (linked by solid line),
illustrating one particular mode of variability in pose.