Digital Signal Processing Reference
In-Depth Information
Fig. 2.10
Extracted results
by different thresholds with
(
a
)
k
1
n
=
1,
k
2
n
=
3,
k
1
s
=
3;
(
b
)
k
1
n
=
2,
k
2
n
=
3,
k
1
s
=
3
2.3
Color Object Segmentation Using Adaptive
Eigen-Subspaces
In the previous section, we have shown that the eigen-subspace is an effective
method for separating the signal and noise components. While segmenting the color
video objects, several difficulties could be met. If the object is even originally with
the same color, it could show distinct properties under different lightening condi-
tions such as shade. In order to solve such problem, an adaptive eigen-subspace seg-
mentation (AESS) algorithm is proposed [
21
]. The proposed method can estimate
and adaptively adjust the eigenvectors under segmentation procedure. Although the
object color is changed with different shade, it is still successfully extracted by using
this algorithm. Accompaning with the proposed AESS algorithm, three searching al-
gorithms are used to effectively and efficiently locate the possible pixel. Both AESS
and the proposed searching algorithms will be discussed in the following sections.
2.3.1
Adaptive Eigenvector Estimation
The eigen-subspace transformation that we have stated in the previous section was
applying the same eigenvectors through the simulation. Sometimes, the eigenvectors
need to be adaptively adjusted according to different simulation conditions. It is
difficult to segment the color object with different shade by just using the same
eigenvectors. We introduce a method that can adaptively adjust the eigenvectors in
simulation. As stated in (
2.1
), we can choose an initial RGB color pixel to form a
vector
s
k
. Then, the related covariance matrix
R
s
can be obtained using (
2.2
). With
covariance matrix,
R
s
, we can get the initial eigenvectors
w
0
of the selected color
pixel. Also, we can transform the initial RGB color pixels
s
0
to eigen-space using
eigenvectors
w
0
as following:
w
0
y
0
=
·
s
0
.
(2.32)
The term
y
0
is projected vectors of the initial color sample, which forms both
the signal space and noise spaces. We can iteratively update the next eigenvectors
with
y
0
. The AESS method is illustrated by the following equations as:
w
k
=
s
k
y
k
.
w
k
±
2
μ
(2.33)
