Image Processing Reference
After the signicant peaks are selected, the valleys are obtained by nding the minimum
values between every two peaks.
Region merging is the nal stage in the process of segmentation. Obtaining clusters on
the basis of peaks and valleys usually results in over-segmentation. Many small regions are
generated and some of the regions may contain very few pixels. Such small regions must be
merged with the closest large regions. The region merging is carried out in two steps. In the
rst step, we check the size of the regions. The regions with number of pixels less than some
predened threshold are merged with the nearest regions. The process is repeated until the
number of pixels in each region is greater than the threshold. Experimentally we found
that threshold 0.1% of the total number of pixels in the image is appropriate. In the second
step, we check for the color similarity between two regions. Two regions are merged if the
the color dierence between these two regions is less than the predened threshold. The
process is repeated until the distance between any two regions in the image is greater than
the predened distance. Here also, experimentally we nd that distance of 20 is appropriate
threshold for region merging.
In this section, we present the experimental results of the roughness index based segmen-
tation algorithm. The qualitative and also the quantitative comparisons of the algorithm
with the histogram based and the histon based segmentations are given here. Several im-
ages obtained from Berkeley database (Martin, Fowlkes, Tal, and Malik, 2001) were used
The Figures 10.3 to 10.8 display the visual comparison of the roughness index based
algorithm with the histogram based segmentation and the histon based segmentation.
In spite of the signicant advancement in image segmentation techniques, it is still an ill-
dened problem as there is no unique ground-truth segmentation of an image against which
the output of an algorithm may be compared (Unnikrishnan, Pantofaru, and Hebert, 2005,
2007). The evaluation of the segmentation algorithms thus far has been largely subjective
and the eectiveness of the algorithm is judged only on intuition and results are given
in the form of few segmented images. To address the problem of quantitative evaluation
of segmentation algorithms several measures proposed in the literature include: Global
consistency Error (GCE), Local Consistency Error (LCE) (Martin, 2002), and Probabilistic
Rand Index (PRI) (Unnikrishnan et al., 2005) and PSNR (Makrogiannis, Economou, and
Fotopoulos, 2005). We used PSNR and PRI for evaluation of the segmentation performance
of our algorithm. The two measures are briey presented here.
The PSNR measure between the image I and the segmented image S is given by
p 255 2 rc
P r i=1 P c j=1 P p k=1 [I (i;j;k) S (i;j;k)] 2
PSNR (I;S) = 10 log 10
where r, c, and p are the number of rows, columns and color components of the image,
respectively. The PSNR represents region homogeneity of the nal partitioning (Makro-
giannis et al., 2005). The higher the value of PSNR the better is the segmentation. The
higher the value of PSNR, the better is the segmentation.