2.2. Unconstrained correlation filter design
The traditiona l correlation filters [21,24–26]in1D-CFAarebased
on the assumption that the correlation peak amplitude should satisfy
aspecified value (i.e., the origin correlation outputs are restricted to
1foraspecific class and 0 for the others). However , the overall
performance of those filters can become worse for unseen patterns if
the correlation peak values are constrained to some specified
constant values during the filterdesign,whichmotivatesustodesign
the filter in the unconstrained form.
UOTF is a traditional unconstrained correlation filter. The
design criterion of UOTF is to: (i) minimize the average energy
and noise variance of the whole correlation plane for all the
samples, and (ii) maximize the origin correlation outputs for the
intra-class samples. However, the minimization of (i) cannot
guarantee that the origin correlation outputs for the extra-class
samples (used to form the feature) are minimal. As a result,
although UOTF [28] tries to overcome the generalization problem
by removing the hard constraints of OTF, UOTF fails in 1D-CFA (see
Section 4 for the experimental results). Therefore, in this paper, we
propose to directly optimize the origin correlation outputs and
take the extra-class samples and intra-class samples into respec-
tive considerations during the filter design.
In the following, we describe the details of the proposed
UOOTF. For the clarity of presentation, vectors are denoted by an
arrow on top of the alphabet. Upper case symbols refer to
quantities in the frequency plane terms, while lower case symbols
represent quantities in the space domain.
1D-CFA designs a correlation filter for each class. Let the filter
trained for the l-th class be h
!
l
, and o
!
l
i
be the output of h
!
l
in
response to y
!
i
.Wehave
o
!
l
i
ðnÞ¼ y
!
i
ðnÞ⊙ h
!
l
ðnÞ; ð1Þ
where ⊙ is a correlation function; y
!
i
is the low-dimensional PCA
feature for the i-th training image; and n is the feature index in the
spatial domain.
Eq. (1) can be expressed in the frequency domain by using the
1D Fourier transform as follows:
o
!
l
i
ðnÞ¼ ∑
p−1
k ¼ 0
Y
!
i
ðkÞ
n
H
!
l
ðkÞe
j2πkn=p
: ð2Þ
here Y
!
i
and H
!
l
represent the 1D Fourier transforms of y
!
i
and h
!
l
,
respectively; p is the reduced dimensionality of the PCA subspace;
k is the feature index in the frequency domain; and ‘
n
’ denotes the
conjugate operator. According to (2), the origin correlation output
(n¼0) is the inner product of the input signal and the correlation
filter in the frequency domain.
The framework of the UOOTF design is shown in Fig. 3. For the
extra-class samples, UOOTF tries to balance the tradeoff between
the origin correlation output energy and the origin correlation
output noise variance. It can be derived by minimizing the
weighted sum of the origin energy j o
!
l
i
ð0Þj
2
and the origin noise
variance j n
!
l
i
ð0Þj
2
for the extra-class samples, which is expressed as
min
H
!
l
ω
s
1
N
el
∑
N
el
i ¼ 1
j o
!
l
i
ð0Þj
2
!
þ ω
n
1
N
el
∑
N
el
i ¼ 1
j n
!
l
i
ð0Þj
2
!
¼ min
H
!
l
ω
s
1
N
el
∑
N
el
i ¼ 1
j Y
!
Eþ
i
H
!
l
j
2
!
þ ω
n
1
N
el
∑
N
el
i ¼ 1
j N
!
Eþ
i
H
!
l
j
2
!
¼ min
H
!
l
ω
s
H
!
lþ
R
l
Y
H
!
l
þ ω
n
H
!
lþ
CH
!
l
; ð3Þ
where R
l
Y
¼ 1=N
el
∑
N
el
i ¼ 1
Y
!
E
i
Y
!
Eþ
i
, and Y
!
E
i
ði ¼ 1; …; N
el
Þ is the 1D
Fourier transform of i-th the extra-class sample for the l-th class.
C ¼ 1=N
el
∑
N
el
i ¼ 1
N
!
E
i
N
!
Eþ
i
, and N
!
E
i
ði ¼ 1; …; N
el
Þ is the 1D Fourier
transform of the i-th extra-class noise sample for the l-th class;
C is usually set as a diagonal matrix whose diagonal elements
represent the noise power spectral density (in fact, C can also be
viewed as a regularization term); ‘+’ represents the conjugate
transpose; ω
s
and ω
n
ð0≤ω
s
; ω
n
≤1Þ are the tradeoff parameters; N is
the number of all the training samples and N
l
is the number of
training samples for the l-th class; N
el
¼ N−N
l
denotes the number
of extra-class training samples for the l-th class.
For the intra-class samples, we try to maximize the average
origin correlation output, which is given by
max
H
!
l
1
N
l
∑
N
l
i ¼ 1
Y
!
Iþ
i
H
!
l
!
¼ max
H
!
l
ðM
!
lþ
H
!
l
Þ; ð4Þ
Fig. 2. Feature extraction in 1D-CFA. Note that FFT is the Fast Fourier Transform
which effectively computes the discrete Fourier transform.
Fig. 3. Framework of the UOOTF design.
Y. Yan et al. / Neurocomputing 119 (2013) 201–21 1 203