Graphics Reference
In-Depth Information
of the discriminative information in DCT coefficients is a key step in order to obtain
good performance of DCT based infrared face recognition method [5].
Partial Least Squares (PLS) is a supervised effective discriminative dimension
reduction technique [11, 12] and has been successfully applied to many vision
applications including face recognition [2, 9, 10]. In this essay, we use PLS to find a
much smaller number of discriminative factors in DCT features. Experimental results
show that PLS further improves the recognition performance based on DCT+LDA
features. This is because PLS basis projects the feature vectors into a latent space in
which feature vectors corresponding to the same subject are closer than the feature
vectors corresponding to different subjects.
2
Discrete Cosine Transformation
The discrete cosine transformation (DCT) is a popular image compression method
[5]. The nuclear transformation of the discrete cosine transformation is the cosine
function of real, thus the calculation complexity of DCT is simple and its information
packing ability closely approaches PCA. Another merit of the DCT is that it can be
implemented efficiently using the Fast Fourier Transform (FFT).
For a M × N digital image
( )
f
x
,
y
Cuv
is shown
(,)
, its two-dimensional DCT,
in the following equation:
MN
−−
11
(2
xu
+
1)
ˀ
(2
yv
+
1)
ˀ
Cuv
(,)
=
auav
()()
f x y
(, ) cos
×
cos
(1)
2
M
2
N
xy
==
00
u
=
0,1,
,
M
−
1;
v
=
0,1,
,
N
−
1
( )
C
,
is the result of DCT which is the DCT coefficient that represents
the purpose of study in this essay. Please be aware that
v
where
a
(
u
)
a
(
v
)
、
are defined
respectively as:
1
M
u
=
0
a
(
u
)
=
2
M
u
=
1
2
,
3
,
,
M
−
1
(2)
1
N
v
=
0
a
(
v
)
=
2
N
v
=
1
2
,
3
,
,
N
−
1
Based on high-compression characteristics and valuable information packing
ability of DCT, it can be used for feature extraction of infrared face recognition to
reduce the relevance of infrared face data [6, 7]. When reconstructing the image using
DCT coefficient, retaining few low-frequency component of DCT and rounding down
mostly high-frequency component, still get the restore images that is similar to the
original images using anti-transformation. The original infrared face, corresponding to
the DCT coefficients and the reconstructed image using 1 / 25 of the DCT coefficients
are shown in Figiure1. As we can see, the majority of important figure (including
nose, mouth, cheeks, etc.) in the restore infrared face is preserved.
Search WWH ::
Custom Search