09-11-2012, 03:22 PM
Two-dimensional Laplacianfaces method for face recognition
1Two-dimensional Laplacianfaces.pdf (Size: 252.4 KB / Downloads: 38)
Abstract
In this paper we propose a two-dimensional (2D) Laplacianfaces method for face recognition. The new algorithm is developed based on
two techniques, i.e., locality preserved embedding and image based projection. The 2D Laplacianfaces method is not only computationally
more efficient but also more accurate than the one-dimensional (1D) Laplacianfaces method in extracting the facial features for human face
authentication. Extensive experiments are performed to test and evaluate the new algorithm using the FERET and the AR face databases. The
experimental results indicate that the 2D Laplacianfaces method significantly outperforms the existing 2D Eigenfaces, the 2D Fisherfaces and
the 1D Laplacianfaces methods under various experimental conditions.
Introduction
The Laplacianfaces method is a recently developed face
recognition method [1]. It is a natural generalization of the locally
linear embedding (LLE) [2] algorithm, which has been
shown to be able to effectively handle the nonlinearity of the
image space for dimensionality reduction. The main idea of the
Laplacianfaces is to find out a low-dimensional representation
of the data that can maximally preserve their locality, i.e., the
pattern of distribution of the data in the local neighborhoods of
the sample space. Differing from the Eigenfaces and the Fisherfaces,
which search for the optimal projections by analyzing
the global patterns of the data density, the Laplacianfaces
method seeks its optimal solutions by examining closely the
local geometry of the training samples. The features learned
are thus quite effective in maintaining the locality of the training
data, making it robust to the outlier samples for training
and suitable for classification with neighborhood based k nearest
neighbor method.
Experimental results
In this section, we experimentally evaluate the proposed
2D Laplacianfaces method on two well-known face databases,
FERET and AR. The FERET database is employed to test the
performance of the face recognition algorithms when various
numbers of samples are selected for training, while the AR
database is used to assess its performance when the face images
are taken with the variations of illuminations, facial expressions
and time sessions. The experiments are performed on
a Pentium 4 2.6 GHz PC with 512MB RAM memory under
Matlab 7.1 platform.
Results on AR database
The AR face database [11] consists of over 4000 face images
of 126 individuals taken in two time sessions under the
variations of illuminations, facial expressions and occlusion
conditions. Each person has 26 images. In our experiment we
consider using a subset of 14 images of each person for training
and testing. Fig. 2 shows the selected sample images of one
subject.
In Fig. 2, the images (a)–(g) and (n)–(t) are drawn from
the first and the second time sessions, respectively. For each
session the first four images (a)–(d) and (n)–(q) involve the
variation of facial expressions (neutral, smile, anger, scream)
while the images (e)–(g) and ®–(t) are taken under different
lighting conditions (left light on, right light on, all sides light
on).
Conclusion
In this paper, we developed the two-dimensional (2D) Laplacianfaces
method and applied it to the face recognition problem.
The proposed method has the following three properties:
First, it can maximally preserve the locality of the geometric
structure of the sample space to extract the most salient features
for classification. The learned local patterns of the training
data are suitable for the neighborhood based kNN queries in the
projected low-dimensional feature space.