Image Processing Reference
In-Depth Information
ods by [Kawase85] (K-T methods), by [Tsao81] (T-F method), and by
[Saito95, Saito96] (S-T method) were used for experiments. Since all these
methods were developed for a binary image, information concerning a den-
sity value is not used for thinning. An artificial image (presented above) was
processed by these algorithms after an input image was binarized by thresh-
olding with a suitable threshold value. A rough tendency was observed in the
experiment using an artificial figure.
In Algorithm 6.2 the extracted center line is very close to an ideal core line
except for in Case 2 of Fig. 6.3. The reason that many short branches appear
in the result is that the local maxima of density value distribution inside a
figure were erroneously regarded as edge points. The effect of the noise in
a density value was suppressed to some extent in Algorithm 6.3 compared
to Algorithm 6.2. This is due to the smoothing effect of GWDT. On the
other hand, the result is affected more by the noise in the shape in Case 3 of
Fig. 6.3. The effect of the shape of an input figure seems to increase by the
use of GWDT.
The K-T method caused degeneration in Object 2, and could not extract
the centerline satisfactorily in the branching part in Object 1 and Case1. The
result was severely affected by shape noise in Case 3.
The T-F method extracted too many plane-like components, in particular
in Case 1 and Case 3. This suggests that the method is affected relatively
more by the rotation of a figure and by the shape noise.
Two branching parts were extracted for Object 1 by the S-T method. In the
results for Object 2 and Case 1, the shapes of branching parts are unnatural.
The effect of shape noise seems to be more severe compared to Algorithm 6.2
and 6.3 in Case 3.
Summarizing these experimental results for an artificial image, Algo-
rithm 6.2 is affected by the density value distribution more than the shape of
a figure. It works effectively in the way that density values are higher in the
vicinity of ideal centerlines, but there is sensitivity to density noise.
A few algorithms have been developed to extract ridgelines of a 2D gray-
tone image [Toriwaki75, Enomoto75, Enomoto76, Naruse77a]. By extending
them to a 3D image, we will be able to derive algorithms with characteristics
similar to Algorithm 6.2, although the topology of a figure is not always
preserved. Algorithm 6.3 will provide a compromise between the centerline of
a figure and a ridgeline of the density value distribution.
The followings is known from experimental results using a real CT image
(Fig. 6.4, Object 3).
The core lines by Algorithm 6.2 and 6.3 are close to an ideal core line.
On the other hand, the core lines by S-T method and K-T method are not
smooth due to unevenness of the surface of regions used as an input figure
in the experiment (= blood vessel regions), extracted automatically from 3D
CT images.
The shape of a branching part is not suitable. Many plane-like components
remain in the results when the T-F method is used. Algorithm 6.2 and 6.3
Search WWH ::
Custom Search