Graphics Reference
In-Depth Information
A P bits binary string is obtained after the process. Then convert the binary string
into a decimal number. Thus, the feature value, namely the LBP value of the center
pixel of a circular neighborhood whose radius is R and the number of neighborhood
pixels is P can be got. The process can be expressed as Eq. 3:
=
P
1
p
LBP
=
s
(
g
g
)
2
(3)
P
,
R
p
c
.
p
0
For the LBP operator model shown in Fig. 1, when the image is rotated, the ar-
rangement of the data of the window will change accordingly. As a consequence, there
are many cases of the LBP values obtained as the preceding steps occurring for a same
image. To solve this problem, Tmeanpaa et. al [14] have improved the original LBP
algorithm, redefining the LBP operator as follows:
{
}
ri
,
LBP
=
min
ROR
(
LBP
(
,
i
))
,
i
=
0
,...,
P
-
1
. (4)
P
R
P
,
R
As seen from Eq. 4, after the image rotated for i times, select the minimum value of
LBP as the neighborhood LBP value. Thus the improved LBP operator is invariable to
rotation and can eliminate the impact during the image rotation.
According to Eq. 3, 2 p kinds of cases can occur after being dealt with the basic LBP
algorithm to an image. To be specific, for P=8 , 2 8 = 256 kinds of cases will be got, and
for P=24 , the kinds of situations will reach to 2 24 . So since, the dimension of LBP
characteristic is quite high and the calculation amount is too large. It does not contri-
bute to conduct follow-up studies. For this situation, Ojala [13] has proposed the concept
of uniform patterns. That is to say, for a local binary pattern string in its circulation
state, if the times of the change between 0 and 1 are no more than twice, then this local
binary pattern is regarded as a uniform pattern. Otherwise it belongs to a non-uniform
pattern. It can be expressed as follows:
=
P
1
s
(
g
g
),
U
2
riu2
,
p
c
LBP
=
P
R
p
0
(5)
P
+
1
otherwise
Where,
=
P
1
U
=
s
(
g
g
)
s
(
g
g
)
+
s
(
g
g
)
s
(
g
g
)
(6)
p
1
c
0
c
p
c
p
1
c
p
1
From Eq. 5 and Eq. 6, the times of 1 appearing in a uniform pattern is considered as
the LBP value of the neighborhood. For the non-uniform patterns, the LBP value of the
neighborhood is set as P +1 . After improvements, the dimension of LBP values drops
to P +2 . The dimension of the feature values is significantly reduced and the efficiency
of the algorithm is effectively improved.
 
Search WWH ::




Custom Search