Pattern Recognition 41 (2008) 3813 -- 3821
Contents lists available at ScienceDirect
Pattern Recognition
journal homepage: www.elsevier.com/locate/pr
Locally linear discriminant embedding: An efficient method for face recognition
Bo Li
a,b
, Chun-Hou Zheng
c
, De-Shuang Huang
a,∗
a
Intelligent Computing Lab, Institute of Intelligent Machine, Chinese Academy of Science, P.O. Box 1130, Hefei, Anhui 230031, China
b
Department of Automation, University of Science and Technology of China, Hefei, Anhui 230027, China
c
College of Information and Communication Technology, Qufu Normal University, Rizhao, Shandong 276826, China
ARTICLE INFO ABSTRACT
Article history:
Received 22 September 2007
Received in revised form 12 April 2008
Accepted 27 May 2008
Keywords:
Feature extraction
Dimensionality reduction
Manifold learning
Locally linear embedding
Face recognition
In this paper an efficient feature extraction method named as locally linear discriminant embedding
(LLDE) is proposed for face recognition. It is well known that a point can be linearly reconstructed by
its neighbors and the reconstruction weights are under the sum-to-one constraint in the classical locally
linear embedding (LLE). So the constrained weights obey an important symmetry: for any particular data
point, they are invariant to rotations, rescalings and translations. The latter two are introduced to the
proposed method to strengthen the classification ability of the original LLE. The data with different class
labels are translated by the corresponding vectors and those belonging to the same class are translated
by the same vector. In order to cluster the data with the same label closer, they are also rescaled to some
extent. So after translation and rescaling, the discriminability of the data will be improved significantly.
The proposed method is compared with some related feature extraction methods such as maximum
margin criterion (MMC), as well as other supervised manifold learning-based approaches, for example
ensemble unified LLE and linear discriminant analysis (En-ULLELDA), locally linear discriminant analysis
(LLDA). Experimental results on Yale and CMU PIE face databases convince us that the proposed method
provides a better representation of the class information and obtains much higher recognition accuracies.
© 2008 Elsevier Ltd. All rights reserved.
1. Introduction
Over the past few decades, face recognition has received a lot of
attention since the wide applications in many fields, such as video
coding, surveillance and human–computer interface, and many
face recognition techniques have been developed. Among them,
appearance-based methods are well studied. Two issues are central
to appearance-based face recognition: one is feature extraction for
face representation; the other is classification of a new face im-
age on the basis of the chosen features. When carrying out these
methods, a face image of size n × m is represented as a vector in
the image space
R
n×m
. However, the n × m-dimensional space is
too large to perform fast face recognition. A widely used way to
attempt to resolve the problem is feature extraction. Being the key
step in appearance-based face recognition, feature extraction aims
to project the input data into a feature space that reflects the inher-
ent structure of the original data and holds the useful information
as much as possible. Thus the low dimensional representations of
the faces can be obtained. Based on these representations, the new
∗
Corresponding author. Tel.: +86 0551 5592751.
E-mail address: dshuang@iim.ac.cn (D.-S. Huang).
0031-3203/$ - see front matter
© 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.patcog.2008.05.027
face images can be easily projected to the low dimensional space. At
last a suitable classifier is adopted to predict the labels of these new
face images. In this study, we shall focus on the topics of feature
extraction for face recognition.
Currently, researchers have developed many feature extraction
techniques for face recognition. These methods can be categorized
into two classes based on either or not taking the class information
into account: supervised or unsupervised. They are also broadly
partitioned into linear methods and nonlinear ones. Linear feature
extraction seeks a meaningful low dimensional subspace in a high
dimensional input space by linear transformation. The subspace can
provide a compact representation of the input data when the struc-
ture of data embedded in the input space is linear. Among all the
linear feature extraction methods, the most well known are prin-
cipal component analysis (PCA) [1] and linear discriminant analysis
(LDA) [2].
PCA projects the original data into a low dimensional space, which
is spanned by the eigenvectors associated with the largest eigenval-
ues of the covariance matrix of all the sample points, where PCA is
the optimal representation of the input data in the sense of mini-
mizing mean squared error (MSE) [1]. However, PCA is completely
unsupervised because of not taking the class information of the in-
put data into account, which may probably discard much useful in-
formation and weaken the recognition accuracy, especially when the
number of sample points is very large [3].