Using Class Specific Linear Projection ...... for each class lies near a linear subspace. .... gitudinal and latitudinal line on the right side of the illustration has a.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 7, JULY 1997
711
Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection Peter N. Belhumeur, Joao P. Hespanha, and David J. Kriegman Abstract—We develop a face recognition algorithm which is insensitive to large variation in lighting direction and facial expression. Taking a pattern classification approach, we consider each pixel in an image as a coordinate in a high-dimensional space. We take advantage of the observation that the images of a particular face, under varying illumination but fixed pose, lie in a 3D linear subspace of the high dimensional image space—if the face is a Lambertian surface without shadowing. However, since faces are not truly Lambertian surfaces and do indeed produce self-shadowing, images will deviate from this linear subspace. Rather than explicitly modeling this deviation, we linearly project the image into a subspace in a manner which discounts those regions of the face with large deviation. Our projection method is based on Fisher’s Linear Discriminant and produces well separated classes in a low-dimensional subspace, even under severe variation in lighting and facial expressions. The Eigenface technique, another method based on linearly projecting the image space to a low dimensional subspace, has similar computational requirements. Yet, extensive experimental results demonstrate that the proposed “Fisherface” method has error rates that are lower than those of the Eigenface technique for tests on the Harvard and Yale Face Databases.
Index Terms—Appearance-based vision, face recognition, illumination invariance, Fisher’s linear discriminant.
1 INTRODUCTION
!
!
#
!
"
$
!
"
$
!
"
4
4
%
$
,
$
)
!
. &
!
!
"
)
5
/
2
6
)
/
+
2
7
$
#
!
"
'
$
(
!
"
8
9
)
$
!
!
!
*
4
;
!
"
.
?
"
+
)
4
$
#
,
$
!
-
&
(
.
(
4
!
$
#
>
?
/
0
!
$
!
(
$
)
+
1
2
"
@
.
A
!
1
!
3
!
$
!
$
&
"
C
B
( 8
!
>
?
4
"
"
%
4
0
"
————————————————
• E
F
K
G
R
[
[
q
h
r
z
r
z
w
s
x
u
s
r
w
L
O
k
x
\
l
y
K
L
U
w
F
J
O
v
y
\
H
J
G
u
{
w
r
U
t
v
I
S
s
s
H
U
{
v
x
N
{
I
s
{
F
L
L
G
O
X
P
Q
n
]
L
O
{
G
X
G
]
Q
^
J
H
S
~
G
U
R
Q
K
_
H
L
G
o
L
P
Q
p
w
O
a
{
K
`
H
|
x
L
U
u
S
G
{
T
G
M
Z
O
}
G
I
J
a
W
J
X
I
X
H
J
b
F
q
O
K
G
G
r
Q
H
N
M
w
U
c
T
V
H
H
Q
~
O
`
F
M
G
H
O
K
Q
p
Q
X
a
H
P
H
{
v
u
{
Q
W
E
U
P
d
G
Z
e
G
W
K
f
I
{
Q
g
J
d
L
h
K
i
U
g
X
Y
e
j
w
z
w
r
x
s
x
r
s
s
}
w
w
{
{
y
{
w
w
x
{
s
s
z
v
u
{
z
x
u
r
w
z
x
v
{
y
t
{
r
{
u
w
{
~
s
}
{
T
J
Z
Z
s
}
G
Z
{
}
w
r
v
z
J
G
~
s
t
F
G
}
r
J
Q
G
|
y
O
S
x
O
]
I
{
s
G
Q
F
v
t
z
L
[
G
r
r
x
U
w
H
H
m
z
M
D
r
x
z
v
{
y
Fig. 1. The same person seen under different lighting conditions can appear dramatically different: In the left image, the dominant light source is nearly head-on; in the right image, the dominant light source is from above and to the right.
0162-8828/97/$10.00 © 1997 IEEE
712
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 7, JULY 1997
!
.
)
5
/
2
A
B
7
%
#
6
)
C
+
!
1
&
"
#
$
"
2
#
:
%
!
3
)
5
?
!
!
!
5
3
#
%
#
8
!
%
"
'
0
!
"
0
!
)
4
!
!
?
!
!
?
)
5
6
@
2
)
2 @
*
.
0
)
!
0
"
#
1
8
"
!
!
!
*
"
&
!
!
$
"
-
)
1
9
:
;
8
.
+
?
;
:
;
?
%
$
"
!
>
5
!
"
9
'
5
:
0
,
"
"
@
!
@
?
4
¶
¸ Ü
Æ
ß
¯
à
á
â
$
#
2.1 Correlation
>
:
°
$
4
BELHUMEUR ET AL.: EIGENFACES VS. FISHERFACES: RECOGNITION USING CLASS SPECIFIC LINEAR PROJECTION
713
Ç
$ Ç
Æ
Ð
Ç
ç
ã
ä
å
1
æ
Æ
)
-
&
!
.
6
è
è
è
ç
µ
¶
é
Ã
í
$
!
/
"
î
!
è
1
2
"
"
©
¬
¨
"
"
ê
ë
ì
Ð
¬
8
#
/
Æ
*
9
:
!
9
;
?
"
æ
!
#
©
¬
³
2
®
3
ê
#
Ð
Ð
è
ê
ê
è
)
/
2
+
.
6
®
4
$
©
¬
ê
æ
"
"
@
'
ï
ð
ñ
ò
®
%
#
#
"
¬
®
&
!
®
7
.
6 Ç
Ç Æ
5
Æ
Ç
6
7
Æ
¼
8
9
ã
*
+
ä
å
ä
!
8
!
!
4
"
Ç
¼
8
$
Ç
Æ
Ð
Ç
:
ä
æ
!
Æ
)
(
>
?
Ç
Ç Æ
Æ
¼
Ð
Ç
8
¼
Ç 8
ä
ä
)
/
2
!
"
!
Ç
5
6
7
$
:
æ
Ç
Æ
Ç
Æ
¼
8
Ð
Ç
¼
8
Ç
"
æ
ä
)
$
³
ä
'
¨
'
'
¬
#
ï
ñ
;
?
9
"
(
"
'
!
: &
$
1
+
7
*
2
!
8
!
²
4
!
$
Ç
!
9
)
:
$
"
*
$
/
"
!
(
8
"
. 8
7
!
!
TABLE 1 COMPARATIVE RECOGNITION ERROR RATES FOR GLASSES/ NO GLASSES RECOGNITION USING THE YALE DATABASE
9
!
!
$
$
$
= %
-
!
!
)
Glasses Recognition Reduced Space
Method
$
?
"
Error Rate (%)
PCA Fisherface
10 1
52.6 5.3
(
0
*
0
!
+
$
7
"
+
)
!
( >
?
!
!
)
,
)
!
!
%
!
!
4
0
!
8
"
)
)
4
¦
¥
¥
-
«
«
©
¥
¤
"
'
4
$
Ç
!
¨
"
$
"
$
%
$
!
4
!
"
"
"
$
4 CONCLUSION
$
1 Ú
)
$
!
"
+
8
.
$
-
ACKNOWLEDGMENTS
(
!
!
!
" '
3
>
!
(
( ,
(
=
+
C
2
7
"
"
"
+
C
+
= >
!
°
*
'
720
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 7, JULY 1997
*
*
+
C
:
)
C
)
+
"
"