(1) First, this paper proposes a modified MFA to generate a
non-sparse original projection matrix for extracting more
effective d iscriminant features. Compared to MFA,
MMFA removes the information in the null space of
the total scatt er matrix, which is not related with the
discriminant ability.
(2) Second, this paper presents a sparse MMFA (SMMFA)
to generate a sparse projection matrix through applying
the linear Bregman iteration to the ℓ
1
-minimization prob-
lem which is related to the original projection matrix
generated by MMFA. The sparsity of the projection ma-
trix makes SMMFA achieve better generalization ability.
It is the first time for MFA-based method to consider
getting a sparse projection matrix.
The remainder of this paper is organized as follows.
Section 2 describes the proposed SMMFA method in detail.
Section 3 discusses the connections of SMMFA with other
related work. Section 4 shows and analyses experimental re-
sults of the presented method. Section 5 concludes the paper.
2SparsemodifiedMFA
In this section, we address the proposed sparse modified MFA
in detail. The framework of SMMFA is given in Fig. 1, where
P is the original projection matrix generated by the modified
MFA which removes the null space in the total scatter
matrix S
t
, V is the sparse projection matrix obtained by apply-
ing the linearized Bregman iteration to P,and⨂ denotes
the operation of matrix multiplication. The training dataset is
used to learn the original projection matrix P. The framework
of SMMFA can be simply described as follows.
First, SMMFA has two adjacency graphs, the intra-class
graph and the inter-class graph to describe the local geometry
structure in the same class and the local discriminant structure
between different classes, respectively, which is similar to
MFA. Different from MFA, SMMFA defines a total scatter
matrix and removes the null space of it to extract more dis-
criminant features, which will be discussed in detail in the
following section 2.1. Second, the linearized Bregman itera-
tion method is used to obtain the sparse solution of SMMFA
by solving the ℓ
1
-minimization problem on the solution
gained before. Finally, the labels of test data could be predict-
ed by the c lassification model trained in the projected
subspace.
2.1 Modified MFA
This section presents the modified MFA. Assume that we have
asetofN samples: x
i
; y
i
ðÞ
fg
N
i¼1
,wherex
i
∈ ℝ
D
, y
i
∈ {1, 2, …,
c} is the label of x
i
, N is the number of samples, D is the
dimension of each sample, and c is the number of classes.
MMFA defines two adjacency graphs, the intra-class graph
and the inter-class graph. The elements of the intra-class graph
F
w
∈ ℝ
N × N
are defined as follows:
F
w
ij
¼
1; if x
i
∈π
þ
K
x
j
or x
j
∈π
þ
K
x
i
ðÞ
0; otherwise
ð1Þ
where π
þ
K
x
j
denotes the set of K homogenous nearest neigh-
bours of x
j
, x
i
∈π
þ
K
x
j
means that x
i
is the homogenou s
nearest neighbour of x
j
,andx
i
and x
j
belong to the same class.
The elements of the inter-class graph F
b
∈ ℝ
N × N
are defined
as follows:
F
b
ij
¼
1; if x
i
∈π
−
K
x
j
or x
j
∈π
−
K
x
i
ðÞ
0; otherwise
ð2Þ
where π
−
K
x
j
denotes the set of K heterogeneous nearest
neighbors of x
j
,andx
i
and x
j
have different class labels.
Similar to MFA, MMFA is to maximize the inter-class
scatter to extract marginal discriminant information between
different classes, and minimize the intra-class scatter to extract
the local similarity information in the same class. In this way,
MMFA could retain the geometry structure of data.
The intra-class scatter is
∑
ij
∥x
i
−x
j
∥
2
F
w
ij
and the inter-class scatter is
∑
ij
∥x
i
−x
j
∥
2
F
b
ij
Fig. 1 Framework of SMMFA, where P is the projection matrix generated by using the modified MFA, V is the sparse projection matrix, and ⨂ denotes
matrix multiplication
Sparse modified marginal fisher analysis for facial expression recognition 2661