Digital Signal Processing Reference
In-Depth Information
With newly found eigenvectors and eigenvalues, we apply (
2.55
)and(
2.56
)to
perform the FCM classification processes.
5. (Optional). Post-processing procedure is optional for smoothing the results from
Step 4. In our experiments, we do not intent to use any post-processing proce-
dures in order to show the inherent classified capability of this algorithm.
2.4.3
Simulation Results
We directly apply the conventional FCM to these four sequences in the RGB
color planes. The simulation results are shown in Fig.
2.23
. Without any threshold
determination, the desired objects obtained by the traditional FCM are better than
those obtained by the PCT method. However, the segmented results still contain
many unwanted noises.
Without any other assistance, Fig.
2.24
shows the simulation results by apply-
ing conventional FCM to the transformed planes, which are performed by the KL
projections. Compared to Fig.
2.23
, some improvements are achieved. The KL pro-
jections obtained from the desired color samples can translate the image data to the
desired working space in more compaction form. It is reasonable to apply the seg-
mentation efforts on the eigen-subspaces. Theoretically, results of Figs.
2.6
and
2.24
should be identical because of the linear transformation between color-space and
eigenspace. The difference of the results shown in these two figures may be due to
initial data distribution and class centers.
For comparison, we also apply OSFCM algorithm [
33
] to the eigen-subspaces.
Figure
2.25
shows the segmented images obtained from the OSFCM method.
Although all the main objects can be detected, the objects with near color are also
Fig. 2.23
Segmented images
obtained by the conventional
FCM directly applying on
R,G,B planes: (
a
) Mosaic;
(
b
) Ball; (
c
) Akiyo; (
d
)News
sequences
