Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 234

We noted that the image monroe provides also a good result (with only one basin
more) with the parameters a=0.75, b=2 and T=1.
Table 4 and Fig. 18 show the four images of the type “neurone”, where only two
images have the same parameters. One reason is that image “neu4” contains larger
objects than the other neurone images. So image neu4 must not get the same parame-
ters, because the Vincent-Soille algorithm is not invariant with respect to scaling.
Image neu2 differs from neu1 and neu3 by more connected objects.
We also tested a simple pre-processing method, which reduced well the overseg-
mentation of many images.
By our preprocessing method we eliminated some small edges in the gradient im-
age, by setting all pixels having a grey value no greater than a threshold to zero. The
threshold must be quite small (compare Fig. 19 and Fig. 20), eliminating only edges
caused by background noise and not too many significant edges in the image.
A problem related to the determination of the segmentation parameters for the
CBR-based watershed algorithm is how to judge the best segmentation quality.
Evaluation done by humans is subjective and can result in different segmented images
of the same input image. This is of importance if we only want to separate the parrot
(see Fig. 23 a) from the background (Fig. 23 b), or segment parts of the parrot as well

(2, 1, 1) BS=78 (2, 1, 1) BS= 65 (2, 1, 1) BS=67

(a) Image neu3 (b) Result for th =2 (c) Result for th =3 (d) Result for th =4

Fig. 19. Different segmentation results by elimination of smaller edges in the gradient image
with threshold th

(0.75, 2, 1) BS=142 (0.75, 2, 1) BS=140 (0.75, 2, 1) BS=129

(a) Image neu3 (b) Result for th =2 (c) Result for th =3 (d) Result for th =4

Fig. 20. Different segmentation results by elimination of smaller edges in the gradient image
with threshold th