Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 237

Images, which are classified as being similar based on the image features, should
be segmented with the same parameters of the segmentation algorithm, which com-
putes then the best segmentation for any of them.
To understand these properties we use hierarchical clustering. This gives us a
graphical representation of the different image groups. Single linkage is used to show
outliers, while the distance between two classes is defined as the minimal distance.
The question is: What are the right image properties that allow us to map the im-
ages to the right image segmentation parameters for the Watershed Transform?
The image description should reflect the behavioral approach of the Watershed
Transformation and the particular image characteristics, respectively. Therefore we
studied the theoretical details and the implementation limits of the Watershed Trans-
formation in section 4 to get insights into this question. We came to the conclusion
that the distribution of the regional minima is an important criterion for the behavior
of the algorithm.
From this work we decided to test four image descriptions described in the follow-
ing Sections:

• Statistical Feature Image Description,

• Image Description as Marginal Distribution for Column and Lines,

• Image Description by Similarity between the Regional Minima of two Images, and

• Image Description by Central Moments.

5.1 Statistical and Texture Feature Image Description

According to Perner [3], who used this description for meta-learning of the segmenta-
tion parameter for a case-based image segmentation model, we applied statistical
features like centroid, energy, entropy, kurtosis, mean, skewness, variance and varia-
tion coefficient and texture features (energy, correlation, homogeneity, contrast) for
case description. The input image is the gradient image of the original image since the
Watershed Transformation works on that image. First results on this image descrip-
tion are reported in Frucci et al.[10]. The texture feature has been chosen to describe
the particular distribution of the regional minima in an image, while the statistical
features describe the signal characteristics.
Like Perner [3], we determined the distance between two images A and B as
follows:

CC −

1

k

∑

dist = ω

iA iB

,

(6)

AB i

k C − C i = 1

i max i min

C are the
maximum and minimum value of all the images for the i th features in the data base,

C

where k is the number of properties in the data base, and

max min

C

is the value of the i th feature of image A (analogous for B) and the ω are

iA i

weights with