Graphics Reference
In-Depth Information
calculated by quantifying pixels of its neighborhood into a string of binary code. Gen-
erally, the size of neighborhood is defined to 3
×
3. The basic LBP code of the central
pixel is computed as
7
=
LBP
p
=
2
i
b
(
g
(
p
)
−
g
(
p
))
(
)
c
i
c
(1)
i
0
,
where
p
c
is the central pixel and
p
i
is the neighbor pixel of
p
c
,
g
(
p
) is the gray value
of
p
and
b
(
u
) is the binary function that
b
(
u
)=1 if
u
is greater or equal to 0;
b
(
u
)=0
otherwise. The process of LBP is given in Fig. 1.
Fig. 1.
The process of LBP
It can be found that the image is represented by a 256-dimensional (8-bits binary
codes stand for 256 numbers in total) histogram by counting the pixel number of each
LBP code. However, the 256-dimensional representation is redundancy and only 58
codes reflect the primitive structural information such as edges and corners. Then the
dimension of the histogram is usually reduced to 59 by assigning non-uniform pat-
terns to single bin [11].
It is obvious that LBP histogram or its uniform pattern is a global texture descriptor. To
further discover the local structures, we divide the entire image into the non-overlapping
blocks (16
×
16) and then compute uniform pattern of LBP in each block. Finally, pathol-
ogy images can be represented by a sequence of 59-dimensional LBP histograms.
3
High-Level Semantic Mining
3.1
Probabilistic Latent Semantic Analysis (pLSA)
Although block LBP features mentioned above can characterize the pathology im-
ages, they are likely to ignore the high-level semantics that may exist in the image. As
the high-level semantic model, Bag-of-words (BoW) [12] performs
k
-means cluster-
ing on the local feature descriptors to generate the codebook composed of visual
words, and then quantizes the descriptors into the words through nearest neighbor
principal. Finally the image can be represented by words. However, there are syn-
onyms among visual words, which may cause that the semantics of images are not well
revealed. As a well-known topic model, probabilistic latent semantic analysis (pLSA)
[9] model aims to describe the image content by the latent topic-level representation
Search WWH ::
Custom Search