没有合适的资源?快使用搜索试试~ 我知道了~
人脸识别论文
4星 · 超过85%的资源 需积分: 0 658 下载量 102 浏览量
2008-03-25
14:58:13
上传
评论 4
收藏 2.31MB PDF 举报
温馨提示
试读
12页
最新的IEEE上的人脸识别论文
资源详情
资源评论
资源推荐
Eigenfeature Regularization and Extraction
in Face Recognition
Xudong Jiang, Senior Member, IEEE, Bappaditya Mandal, and Alex Kot, Fellow, IEEE
Abstract—This work proposes a subspace approach that regularizes and extracts eigenfeatures from the face image. Eigenspace of the
within-class scatter matrix is decomposed into three subspaces: a reliable subspace spanned mainly by the facial variation, an unstable
subspace due to noise and finite number of training samples, and a null subspace. Eigenfeatures are regularized differently in these three
subspaces based on an eigenspectrum model to alleviate problems of instability, overfitting, or poor generalization. This also enables the
discriminant evaluation performed in the whole space. Feature extraction or dimensionality reduction occurs only at the final stage after
the discriminant assessment. These efforts facilitate a discriminative and a stable low-dimensional feature representation of the face
image. Experiments comparing the proposed approach with some other popular subspace methods on the FERET, ORL, AR, and GT
databases show that our method consistently outperforms others.
Index Terms—Face recognition, linear discriminant analysis, regularization, feature extraction, subspace methods.
Ç
1INTRODUCTION
F
ACE recognition has attracted many researchers in the area
of pattern recognition, machine learning, and computer
vision because of its immense application potential. Numer-
ous methods have been proposed in the last two decades [1],
[2]. However, there are still substantial challenging problems,
which remain to be unsolved. One of the critical issues is how
to extract discriminative and stable features for classification.
Linear subspace analysis has been extensively studied and
becomes a popular feature extraction method since the
principal component analysis (PCA) [3], Bayesian maximum
likelihood (BML) [4], [5], [6], and linear discriminant analysis
(LDA) [7], [8] were introduced into face recognition. A
theoretical analysis showed that a low-dimensional linear
subspace could capture the set of images of an object
produced by a variety of lighting conditions [9]. The agree-
able properties of the linear subspace analysis and its
promising performance achieved in the face recognition
encourage researchers to extend it to higher order statistics
[10], [11], nonlinear methods [12], [13], [14], Gabor features
[15], [16], and localit y preserving projections [17], [18].
However, the basic linear subspace analysis has still out-
standing challenging problems when applied to the face
recognition due to the high dimensionality of face images and
the finite number of training samples in practice.
PCA maximizes the variances of the extracted features
and, hence, minimizes the reconstruction error and removes
noise residing in the discarded dimensi ons. The best
representation of data may not perform well from the
classification point of view because the total scatter matrix
is contributed by both the within and between-class varia-
tions. To differentiate face images of one person from those of
the others, the discrimination of the features is the most
important. LDA is an efficient way to extract the discrimina-
tive features as it handles the within and between-class
variations separately. However, this method needs the
inverse of the within-class scatter matrix. This is problematic
in many practical face recognition tasks because the dimen-
sionality of the face image is usually very high compared to
the number of available training samples and, hence, the
within-class scatter matrix is often singular.
Numerous methods have been proposed to solve this
problem in the last decade. A popular approach called
Fisherface (FLDA) [19] applies PCA first for dimensionality
reduction so as to make the within-class scatter matrix
nonsingular before the application of LDA. However,
applying PCA for dimensionality reduction may result in
the loss of discriminative information [20], [21], [22]. Direct
LDA (DLDA) method [23], [24] removes null space of the
between-class scatter matrix and extracts the eigenvectors
corresponding to the smallest eigenvalues of the within-class
scatter matrix. It is an open question of how to scale the
extracted features, as the smallest eigenvalues are very
sensitive to noise. The null space (NLDA) approaches [25],
[26], [22] assume that the null space contains the most
discriminative information. Interestingly, this appears to be
contradicting the popular FLDA that only uses the principal
space and discards the null space. A common problem of all
these approaches is that they all lose some discriminative
information, either in the principal or in the null space.
In fact, the discriminative information resides in both
subspaces. To use both subspaces, a modified LDA approach
[27] replaces the within-class scatter matrix by the total scatter
matrix. Subsequent work [28] extracts features separately
from the principal and null spaces of the within-class scatter
matrix. However, the extracted features may not be properly
scaled andundueemphasis isplaced on the null space in these
two approaches due to the replacement of the within-class
scatter matrix by the total scatter matrix. The dual-space
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 30, NO. 3, MARCH 2008 383
. The authors are with the School of Electrical and Electronic Engineering,
Nanyang Technological University, Nanyang Link, Singapore 639798.
E-mail: {exdjiang, bapp0001, eackot}@ntu.edu.sg.
Manuscript received 27 June 2006; revised 6 Nov. 2006; accepted 15 May
2007; published online 31 May 2007.
Recommended for acceptance by T. Tan.
For information on obtaining reprints of this article, please send e-mail to:
tpami@computer.org, and reference IEEECS Log Number TPAMI-0475-0606.
Digital Object Identifier no. 10.1109/TPAMI.2007.70708.
0162-8828/08/$25.00 ß 2008 IEEE Published by the IEEE Computer Society
approach (DSL) [21] scales features in the complementary
subspace by the average eigenvalue of the within-class scatter
matrix over this subspace. As eigenvalues in this subspace are
not well estimated [21], their average may not be a good
scaling factor relative to those in the principal subspace.
Features extracted from the two complementary subspaces
are properly fused by using summed normalized distance
[14]. Open questions of these approaches are how to divide
the spaceinto the principal and the complementary subspaces
and how to apportion a given number of features to the two
subspaces. Furthermore, as the discriminative information
resides in both subspaces, it is inefficient or only suboptimal
to extract features separately from the two subspaces.
The above approaches focus on the problem of singularity
of the within-class scatter matrix. In fact, the instability and
noise disturbance of the small eigenvalues cause great
problems when the inverse of matrix is applied such as in
the Mahalanobis distance, in the BML estimation, and in the
whitening process of various LDA approaches. Problems of
the noise disturbance were addressed in [29], and a unified
framework of subspace methods (UFS) was proposed. Good
recognition performance of this framework shown in [29]
verifies the importance of noise suppression. However, this
approach applies three stages of subspace decompositions
sequentially on the face training data, and dimensionality
reduction occurs at the very first stage. As addressed in the
literature [20], [21], [22], applying PCA for dimensionality
reduction may result in the lost of discriminative information.
Another open question of UFS is how to choose the number of
principal dimensions for the first two stages of subspace
decompositions before selecting the final number of features
at the third stage. The experimental results in [29] show that
recognition performance is sensitive to these choices at
different stages.
In this paper, we present a new approach for facial
eigenfeature regularization and extraction. Im age space
spanned by the eigenvectors of the within-class scatter matrix
is decomposed into three subspaces. Eigenfeatures are
regulariz ed differently in these subspaces based on an
eigenspectrum model. This alleviates the problem of unreli-
able small and zero eigenvalues caused by noise and the
limited number of training samples. It also enables discrimi-
nant evaluation to be performed in the full dimension of the
image data. Feature extraction or dimensionality reduction
occurs only at the final stage after the discriminant assess-
ment. In Section 2, we model the eigenspectrum, study the
effect of the unreliable small eigenvalues on the feature
weighting, and decompose the eigenspace into face, noise,
and null subspaces. Eigenfeature regularization and extrac-
tion are presented in Section 3. Analysis of the proposed
approach and comparison with other relevant methods are
provided in Section 4. Experimental results are presented in
Section 5 before drawing conclusions in Section 6.
2EIGENSPECTRUM MODELING AND SUBSPACE
DECOMPOSITION
Given a set of properly normalized w-by-h face images, we
can form a training set of column image vectors fX
ij
g, where
X
ij
2 IR
n¼wh
, by lexicographic ordering the pixel elements of
image j of person i. Let the training set contain p persons and
q
i
sample images for person i. The number of total training
sample is l ¼
P
p
i¼1
q
i
. For face recognition, each person is a
class with prior probability of c
i
. The within-class scatter
matrix is defined by
S
w
¼
X
p
i¼1
c
i
q
i
X
q
i
j¼1
ðX
ij
X
i
ÞðX
ij
X
i
Þ
T
; ð1Þ
where
X
i
¼
1
q
i
P
q
i
j¼1
X
ij
. The between-class scatter matrix S
b
and the total (mixture) scatter matrix S
t
are defined by
S
b
¼
X
p
i¼1
c
i
ðX
i
XÞðX
i
XÞ
T
; ð2Þ
S
t
¼
X
p
i¼1
c
i
q
i
X
q
i
j¼1
ðX
ij
XÞðX
ij
XÞ
T
; ð3Þ
where
X ¼
P
p
i¼1
c
i
X
i
. If all classes have equal pri or
probability, then c
i
¼ 1=p.
Let S
g
, g 2ft; w; bg represent one of the above scatter
matrices. If we regard the elements of the image vector and
the class mean vector as features, these preliminary features
will be decorrelated by solving the eigenvalue problem
g
¼
g
T
S
g
g
; ð4Þ
where
g
¼½
g
1
; ...;
g
n
is the eigenvector matrix of S
g
, and
g
is the diagonal matrix of eigenvalues
g
1
; ...;
g
n
corresponding to the eigenvectors.
Suppose that the eigenvalues are sorted in descending
order
g
1
; ...;
g
n
. The plot of eigenvalues
g
k
against the
index k is called eigenspectrum of the face training data. It
plays a critical role in subspace methods as the eigenvalues
are used to scale and extract features. We first model the
eigenspectrum to show its problems in feature scaling and
extraction.
2.1 Eigenspectrum Modeling
If we regard X
ij
as samples of a random variable vector X, the
eigenvalue
g
k
is a variance estimate of X projected on the
eigenvector
g
k
estimated from the training samples. It usually
deviates from the true variance of the projected random
vector X due to the finite number of training samples. Thus,
we model the eigenspectrum in the range subspace
g
k
,
1 k r, as the sum of the true variance component v
F
k
and
a deviation component
k
.Forsimplicity,wecallv
F
k
face component and
k
noise component. As the face
component typically decays rapidly and stabilizes, we can
model it by a function of the form 1 =f that can well fit to the
decaying nature of the eigenspectrum. The function form 1=f
was used in [4] to extrapolate eigenvalues in the null space for
computing the average eigenvalue over a subspace. The noise
component
k
that includes the effect of the finite number of
training samples can be negative if the face component v
F
k
is
modeled by a 1=f function that always has positive values.
We propose to model the eigenvalues first in descending
order of the face component v
F
k
by
^
F
k
¼ v
F
k
þ
k
¼
k þ
þ
k
; 1 k r; ð5Þ
where and are two constants that will be given in
Section 3.1. The modeled eigenspectrum
^
g
k
is then obtained
by sorting
^
F
k
in descending order. As the eigenspectrum
384 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 30, NO. 3, MARCH 2008
decays very fast, we plot the square roots
g
k
¼
ffiffiffiffi
g
k
p
and ^
g
k
¼
ffiffiffiffi
^
g
k
p
for clearer illustration (we still call them eigenspectrum
for simplicity). A typical real eigenspectrum
g
k
and its
model ^
g
k
are shown in Fig. 1, where
k
is a random number
evenly distributed in a small range of ½
g
1
=1000;
g
1
=1000.
In Fig. 1, we see that the proposed model closely
approximates to the real eigenspectrum.
2.2 Problems of Feature Scaling and Extraction
The PCA approach with euclidean distance selects d leading
eigenvectors and discards the others. This can be seen as
weighting the eigenfeatures with a step function as
u
d
k
¼
1;k d
0;k>d:
ð6Þ
This weighting function of PCA w
P
k
¼ u
d
k
is shown in Fig. 1.
In a practical face recognition problem, the recognition
performance usually improves with the increase of d. The
PCA with Mahalanobis (PCAM) distance can be seen as the
PCA with euclidean distance and a weighting function
w
M
k
¼ u
d
k
=
g
k
. This weighting function is shown in Fig. 1.
Using the inverse of the square root of the eigenvalue to weigh
the eigenfeature makes a large difference in the face
recognition performance. The recognition accuracy usually
increases sharply with the increase of d and is better than the
euclidean distance for smaller d. However, the recognition
accuracy decreases also sharply and is much worse than the
euclidean distance for larger d. The BML approach that
weights the principal features ð1 k dÞ by the inverse of
the square root of the eigenvalue also suffers from the
performance decrease with the increase of the d after d reaches
a small value.
Noise disturbance and poor estimates of small eigenva-
lues due to the finite number of training samples are the
culprits. The limited number of training samples results in
very small eigenvalues in some dimensions that may not
well represent the true variance in these dimensions. This
may result in serious problems if their inverses are used to
weight the eigenfeatures. The characteristics of the eigen-
spectrum and the generalization deterioration caused by the
small eigenvalues were well addressed in [30]. To exclude
the small eigenvalues in the discriminant evaluation,
probabilistic reasoning models and enhanced FLDA models
were proposed and compared in [30]. The modeled
eigenspectrum in Fig. 1 that approximates closely to the
real one is resorted in descending order of the face
component v
F
k
and is plotted in Fig. 2. The small noise
disturbances are now visible in Fig. 2 given that the
variances of the face component should always decay.
These small disturbances cause large vibrations of the
inverse eigenspectrum, as shown in Fig. 2. Eigenfeatures of
larger index k are heavily weighted by the scaling factors
that are highly sensitive to noise and training data. This
causes the deterioration of the recognition performance,
especially on the independent testing data.
2.3 Subspace Decomposition
As shown in Fig. 2, small noise disturbances that have little
effect on the initial portion of the eigenspectrum cause large
vibrations of the inverse eigenspectrum in the region of small
eigenvalues. Therefore, we propose to decompose the eigen-
space IR
n
spanned by eigenvectors f
g
k
g
n
k¼1
into three
subspaces: a reliable face variation dominating subspace (or
simply face space) F ¼f
g
k
g
m
k¼1
, an unstable noise variation
dominating subspace (or simply noise space) N ¼f
g
k
g
r
k¼mþ1
and a null space ;¼f
g
k
g
n
k¼rþ1
, as illustrated in Fig. 2. The
purpose of this decomposition is to modify or regularize the
unreliable eigenvalues for better generalization.
JIANG ET AL.: EIGENFEATURE REGULARIZATION AND EXTRACTION IN FACE RECOGNITION 385
Fig. 1. A typical real eigenspectrum, (a) its model and feature weighting/extraction of PCA. The first 50 real eigenvalues and (b) the values of the
model.
Fig. 2. A typical real eigenspectrum, its model sorted in descending
order of the face component and their inverse; decomposition of the
eigenspace into face, noise, and null-subspaces.
剩余11页未读,继续阅读
nacy123
- 粉丝: 0
- 资源: 1
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功
评论9