Image Thresholding using Generalized Rough Sets - Rough Fuzzy Image Analysis: Foundations and Methodologies - page 61

As we have considered the evaluation of bilevel thresholding based segmentation here,
every image under consideration would be separated into two regions. However, the number
of regions in the human labeled segmentation ground truths of the 100 images considered
is always more than two. Now, the LCE measure penalizes an algorithm only if both S H
and S A are not refinements of each other at a pixel and it does not penalize an algorithm if
any one of them is the refinement of the other at a pixel (Martin et al., 2001). Therefore,
the use of LCE measure is desirable in our experiments, as we do not want to penalize an
algorithm when S A is not a refinement of S H at a pixel and S H is a refinement of S A at that
pixel. This aforesaid case is a very highly probable one in our experiments, as the number
regions associated with S A would be much less than S H .

(a) Box plots for algorithm (i) (b) Box plots diagrams for

algorithm (ii)

(c) Box plots for algorithm (iii) (d) Box plots for algorithm (iv)

(e) Box plots for algorithm (v) (f) Box plots for algorithm (vi) (g) Box plots for algorithm (vii)

FIGURE 3.12: Box plot based summarization of segmentation performance by various thresh-
olding algorithms