Information Technology Reference
In-Depth Information
where the notation
denotes a neighborhood of
P
points equally sam-
pled on a circle of radius
R
,
g
c
is the gray-level value of the central pixel and
g
p
are the
(
P, R
)
P
gray-level values of the sampled points in the neighborhood.
P
s
(
x
)
is
1
if
x ≥
0
and
0
otherwise. The
LBP
P,R
operator produces
2
dif-
P
different binary patterns that
canbeformedbythe
P
pixels in the neighbor set. It has been shown that
certain bins contain more information than others [14]. These patterns are
called uniform patterns and are obtained if LBP contains at most two bit-
wise transitions from 0 to 1 or vice versa when the binary string is considered
circular. For example, 00011000, 11000001, 00000000 are uniform patterns.
The uniform patterns represent fundamental local primitives, such as edges
or corners, as show in Figure 4. It was also observed that most of the texture
information (90%) is contained in the uniform patterns [14]. The patterns
which have more than two transitions are given a unique label, therefore the
operator, denoted
LBP
2
ferent output values, corresponding to the
P
bins. The standard opera-
tor
LBP
8
,
2
will result in 256 different labels while
LBP
u
2
P,R
, will have less than
2
u
2
P,R
will have only 59
labels. After the computation of the LBP labels, a histogram is constructed
as follows
x,y
I
(
f
l
(
x, y
)=
i
)
,i
=0
,
1
, ..., L −
1
H
i
=
(3)
where
L
is the number of different labels produced by the LBP operator,
f
l
is the LBP code of the central pixel
andI(A)is1ifAistrueand0oth-
erwise. Moreover, LBP has been extended for multiresolution analysis [14].
The usage of different values for
P
and
R
permits to realize operators for any
quantization of the angular space and for any spatial resolution. The informa-
tion provided by multiple operators is then combined. As the LBP histogram
contains information about the distribution of local micro-patterns over the
whole image, the so computed descriptor represents a statistical description
of image characteristics. This descriptor has been proved to be successful, to-
gether with its original design as texture description and recognition [14], also
in face detection and recognition [1], image retrieval [19] and facial expression
analysis and recognition [17].
Another version of LBP has been recently developed by Heikkila et al.
[8]: the pixels are compared in a different manner, giving a descriptor whose
length is 16 times shorter than LBP. The CS-LBP operator is expressed in
decimal form as
(
x, y
)
(
P/
2)
−
1
p
CS − LBP
P,R
=
s
(
g
p
− g
p
+(
P/
2)
)2
(4)
p
=0
In their work, CS-LBP descriptor is proposed to be a powerful descriptor
as it combines the strengths of the SIFT descriptor and the LBP texture
operator: the method uses a SIFT-like grid approach and replaces SIFT's
Search WWH ::
Custom Search