Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 231

some cases the possibility to check adjacent basins and, hence, to apply flooding or
digging. For an example of thick watershed lines, see Fig. 15, where LO XY is equal to
12, the basins X and Y are those respectively characterized by the local minima equal
to 8, and 6.
If, instead of flooding, Table 1 had suggested to apply digging, we would have ob-
tained mainly the same result for the image in Fig. 15 (b). The only difference would
have been that pixels labeled 12 would have been labeled 8. In that case one obtaines
the parameters At=Dt=8 if the surrounding image of this cutout shown in Fig. 15 is
suitable.
Furthermore, it should be noted that digging over O is not always the best way.
Other minimal criteria for path digging are explained by Bleau and Leon in [16].
In our study, we are interested in analyzing the image properties in order to detect
the proper values for the constants a , b and T . The constants a and b control the influ-
ence of the similarity parameter and the depth. T can be regarded as a threshold. If T
is 0 and

ab , ≥ ,

then we obviously get the same segmentation like the one produced
by the Vincent-Soille algorithm. For

T = 1 ab , = 2

and we often obtain similar

results to those achieved by using the algorithm [17].
Since we have used the Vincent-Soille Watershed Transformation, which is not invari-
ant for image rotation and scaling, in our implementation of algorithms [10] and [11], the
best value of a, b and T can be different for two images after scaling or rotation.
To decide which watershed-based segmentation algorithm to use, we have also
considered the algorithm described in [10] in its original implementation. Making a
compromise between computation costs and quality of the obtained segmentation
results, we have finally opted for the Vincent-Soille algorithm as a basis for the
CBR-based watershed algorithm. In future work, we will carry out further tests on
the different behavior of the basic watershed algorithms. The hope is that the choice
of the basic algorithm can be included as a parameter into a CBR-based watershed
algorithm.

4 Test Images and Parameters for Watershed Segmentation

For our study we used a data base comprised of different types of images, such as e.g.
landscape, medical and biological images, and face images. Nine of these images are
shown in Fig. 16a-j and are used for demonstration purposes of the results throughout
this chapter. The neurone images {neu1;neu2;neu3;neu4;neu4-r180} seem to belong
to the same class, except that neu4_r180 is the 180 degree rotated image of neu4 and
neu4 seems to be a rescaled cut-out of one of the image neurons.
The parameters for the Watershed segmentation were obtained by running the
CBR-based Watershed segmentation and adjusting the parameters until the result had
the best segmentation quality. The resulting parameters are shown in Table 2.