Image Processing Reference
In-Depth Information
where
Hist
p
and
Hist
q
i
are the normalized histograms of
p
and
q
i
respectively, and
the higher Bhattacharyya coefficient indicates the higher similarity between them.
RCA scheme can be formed as follows:
M
t
while
(
φ
ↂ
)
{
M
t
+
1
M
t
+
1
M
t
M
t
=
,
=
M
t
∀
p
∈
,∀
q
i
∈
N
p
(9.9)
θ
(
,
)=
i
=
1
,
2
,...,
k
θ
(
,
)
if
p
q
i
max
p
N
p
M
t
+
1
M
t
M
t
+
1
M
t
=
∪
q
i
,
=
\
q
i
endi f
}
where
M
t
and
M
t
are the sets of marker regions and non-marker regions (labeled as
0 and 1) at the
t
time respectively. The iteration stops when the entire
M
shrinks to
empty set. The RCA rule is simple and required less user inputs. This fact makes it
is conveniently to resolve complex tasks and RCA is essentially a game of life and
easily extended to N-D space [15].
Similar to MSRM, RCA is adaptive to image content and it does not need to set
the regional similarity threshold in advance. Other color spaces, initial segmentation
and distance metrics can also be used in RCA. If the initial segmentation can pro-
vide a good basis, RCA may get better results. Different from MSRM, RCA has the
ability to deal with multi-class problem according to the number of cell states, which
grants it with the flexibility in parallel implementation. Figure 9.3 shows examples
to extract single and multiple objects from background. Although the contour of the
bird is complex and the skin of polar bears is somewhat similar to the snow back-
ground, RCA still successfully separates objects from the background with simple
markers. Other color space, initial segmentation and distance metrics can also be
used in RCA. If the initial segmentation can provide a good basis, RCA may get
better results. Different from MSRM, RCA completes regions merging not consid-
ering the marker or non-marker state, which grants it with the greater flexibility
in parallel implementation. Moreover, RCA has the ability to deal with multi-class
problem according to the number of cell states.
In validation, examples are presented to verify the performance of RCA imple-
mented on the test images woman, church and fish with different color and shape
features respectively. As shown in Figure 9.4, although the objects are complex
and there is no explicit user input background marker, RCA method can still extract
desired objects accurately. In general, RCA algorithm could reliably extract the ob-
ject contour if the markers cover main features of object or background. However, it
may fail when shadow, low-contrast edges and ambiguous areas occur. For example,
a parts of the object regions in fish image are very similar to background. Although
many markers are used, but the result is not satisfied. Figure 9.5 illustrates the dif-
ference between RCA and MSRM segmentation highlighted by red color in the fish
Search WWH ::
Custom Search