Image Processing Reference
In-Depth Information
T4 T7 T5 T1
T2 T1 T4 T6
T4 T3 T2 T7
T2 T6 T5 T3
Fig. 16.10. On the
left
the found texture groups are shown in color, i.e., points sharing the
same color belong to the same group. On the
right
, the boundaries and class labels implied by
this grouping are superimposed the original image
initialization of the clustering, which was random, as well as the required classes,
which was varied from 7 to 9. Figure 16.9 illustrates such a result, the boundaries
and the texture classes of which agree fairly well with the ground truth, shown in
Fig. 16.8.
16.8 Further Reading
A word of caution is in place at this point. A disadvantage of the simple structure of a
resolution pyramid, is its incapacity to keep track of small regions while building up
the pyramid. In other words, this scheme assumes that the regions to be segmented
do not have extreme size variations because very small regions will be considered
noise if one attempts to reduce the noise within large regions by using a too high
L
. On the other hand, a more adequate solution would need general automatic scale
selection, an issue which has proven to be a significant challenge [151].
In the vast clustering literature available for general feature vectors, it can be
noted that different criteria might lead to different results depending if the shape of
the metrics is not reasonable as compared to cluster shapes. It is clear that no cluster-
ing criterion or measure of dissimilarity is universally applicable. For instance, the
single link method is more suitable for elongated clusters while WGSS (within-group
sum of squares) criterion gives good results for ball-shaped structures [87].However,
it can be argued that clustering used to achieve perceptual grouping or image seg-
mentation should not critically depend on the choice of a clustering technique if
features are sufficiently strong to characterize groups. If they are not, it is often more
worthwhile to devote efforts to redesign the features by using human experts than to
