Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 241

Hi the i th
row in the histogram H (analogous for H ). The problem of this measure is that it
cannot handle functional values of zero in the histogram. Then the multiplicand in
formula (9) is not defined. A functional value of zero means that for a column or a
row no foreground pixel exists, which can always happen in a binarized image.
Therefore, although this measure is often cited in the literature as the preferred meas-
ure when comparing two histograms, for the purpose proposed in this paper we can-
not use this measure.
We used different distance functions, like (squared) Euclidean Distance, city block
or Chebyshev Distance to compare two histograms.
The dendrograms of the single-linkage method in Figures Fig. 28 to Fig. 31 show
the results of the marginal distribution using the Euclidean Distance, where Fig. 28
(and Fig. 29 and Fig. 31, respectively) shows the results comparing the column histo-
gram (and row and diagonal histograms, respectively) and Fig. 30 shows the results
comparing the column and row histograms together.

H and H are two histograms, n the size of the diagrams and 1 ()

where

Fig. 28. Dendrogram for image description. Marginal distribution for column using the similar-
ity measure Euclidean Distance

We virtually cut the dendrogram in Fig. 28 by the cophenetic-similarity value
equal to 0.455. However, with this cophenetic-similarity value, images for which we
got different best segmentation parameters were also clustered in different cases.
Therefore the cut-off will be by the cophenetic-similarity value equal to 0.389, be-
cause monroe and neu3 have different best segmentation parameters. By this choice