A Convergent Solution to Matrix Bidirectional Projection
3. In our method, the manifold structure of the
image space, which is modeled by an adja-
cency graph, is explicitly taken into account.
4. Our method can automatically select suit-
able feature dimensionality with which algo-
rithm can obtain comparable recognition per-
formance. This is very important in prac-
tice. The previous methods need to consider
all the possible dimensionality to obtain the
top recognition performance. This is very
time-consuming and inapplicable in real face
recognition system.
2. Laplacian scatter matrix
Let matrix x represent an image with m ×n pixels,
then feature matrix y of image x can be obtained by:
y = U
T
xV (
1)
where U and V are m ×m
′
(m
′
6 m) column projec-
tion matrix and n×n
′
(n
′
6 n) row projection matrix,
respectively.
Suppose we are given N training im-
ages X = [x
1
,x
2
,... ,x
N
] = [X
1
,. .. ,X
i
,. .. ,X
c
] =
[x
(1)
1
,. .. ,x
(i)
j
,. .. ,x
(c)
N
c
] which belong to c different
classes, the ith class has N
i
images(
∑
c
i=1
N
i
= N)
a
nd matrix X
i
= [x
(i)
1
,x
(i)
2
,. .. ,x
(i)
N
i
] consists of the im-
age matrices from the ith class. By representing
each image matrix as an m-set of row vectors, the
row total scatter matrix can be expressed as:
S
row
t
=
1
N
N
∑
i=1
m
∑
j=1
(x
ij
−x
( j)
)
T
(x
ij
−x
( j)
) (2)
=
1
N
N
∑
i=1
(x
i
−x)
T
(x
i
−x) (3)
=
1
2N
2
N
∑
i=1
N
∑
j=1
(x
i
−x
j
)
T
(x
i
−x
j
) (4)
where x, x
ij
and x
( j)
are the mean matrix of all train-
ing images, jth row vector of ith image matrix and
mean vector of jth row vector of all training images,
respectively.
The row within-class scatter matrix can be de-
fined as:
S
row
w
=
1
N
c
∑
i=1
N
i
∑
j=1
(x
(i)
j
−m
i
)
T
(x
(i)
j
−m
i
) (5)
=
1
N
c
∑
i=1
1
2N
i
N
i
∑
j,k=1
(x
(i)
j
−x
(i)
k
)
T
(x
(i)
j
−x
(i)
k
) (6)
where x
(i)
j
and m
i
are the jt
h image matrix of ith class
and mean matrix of ith class, respectively.
The use of manifold information in feature ex-
traction has shown the state-of-the-art face recog-
nition performance
10,23,24
. According to graph em-
bedding theory
25
, we define an undirected weighted
graph G(X,W) to characterize the nonlinear mani-
fold structure of the image set X. The real symmet-
ric matrix W measures similarities of any pairs of
samples. It can be constructed using various simi-
larity criterion, such as Gaussian similarity in Lapla-
cian eigenmap
9
, local neighborhood relationship as
in LLE
11
and also prior class information in super-
vised learning algorithms. Here, the Gaussian simi-
larity is adopted:
w
ij
= exp(−kx
i
−x
j
k
2
/(2
σ
2
)) (7)
Then in order to incorporate the nonlinear mani-
fold structure of face images, we can define the fol-
lowing row total Laplacian scatter:
LS
row
t
=
1
2N
2
N
∑
i, j=1
w
ij
(x
i
−x
j
)
T
(x
i
−x
j
) (8)
=
1
N
2
N
∑
i, j=1
(w
ij
x
T
i
x
i
−w
ij
x
T
i
x
j
) (9)
=
1
N
2
X
′T
(L⊗I
m
)X
′
(10)
where X
′
= [x
T
1
,x
T
2
,. .. ,x
T
N
]
T
, D is a diagonal matrix
with d
ii
=
∑
N
j=1
w
ij
, L = D−W is the Laplacian ma-
trix of graph G, I
m
is identity matrix of order m and
operator ⊗ is the Kronecker product of matrices.
Similarly, in row direction, the image within-
Published by Atlantis Press
Copyright: the authors
865