Graphics Reference
In-Depth Information
Fig. 10.2.
Discrete search space.
E
snake
=
1
0
E
int
(
ν
(
s
)) +
E
image
(
ν
(
s
))
.
(10.1)
Using the discrete notation from Chapter 9, we have
E
int
=
ν
i
+1
−
2
ν
i
+
ν
i−
1
ν
i
+1
−
(10.2)
ν
i−
1
and
2
.
E
image
=
−|∇
I
(
x, y
)
|
(10.3)
The internal energy consists of a thin-plate term only. The relative weight-
ing of the cost components is controlled by a single regularization parameter,
λ
[0
,
1]. By choosing a high value of
λ
, the thin-plate or stiffness term
dominates, which may lead to smooth contours that tend to ignore important
image edges. On the other hand, low values of
λ
allow contours to develop
sharp corners as they attempt to follow all high gradient edges, even those
that may not necessarily be on the desired object's boundary. Once every
contour has been evaluated, the single contour with least cost is the global
solution. The Viterbi algorithm provides a very ecient method to find this
global solution, as described in Section 9.5.
A data set of 19946 Pap stained cervical cell images was available for
testing. The single regularization parameter
λ
was empirically chosen to be
0.7 after trial runs on a small subset of the images. The effect of the choice of
λ
on segmentation accuracy on this trial set is shown by the graph of Figure
10.3. This figure shows a value of
λ
=0
.
7 as being the most suitable for these
particular images. It further shows that acceptable segmentation performance
can be obtained with
λ
ranging from 0.1 to 0.9—an enormous range, which
demonstrates the robustness and suitability of the approach. Every image in
the data set was then segmented at
λ
=0
.
7 and the results verified by eye.
Of the 19946 images, 99.5% were found to be correctly segmented.
With
λ
set at 0.0, the smoothness constraint from the thin-plate term is
completely ignored and the point of greatest gradient is chosen along each of
the search space radii. Previous studies [12] have shown that for approximately
∈
Search WWH ::
Custom Search