Image Processing Reference
In-Depth Information
a
b
c
d
Fig. 5.8 Edge detection applied to an image which has had salt and pepper noise added. a)
Sobel, b) CA rules learnt from training set in figure 5.2, c) CA rules learnt from a version of
the training set in figure 5.2 with added salt and pepper noise, d) CA rules learnt from noisy
data applied to both the image and an inverted version.
Fig. 5.9 Edges detected using Wongthanavasu and Lursinsap's [45] method (upper row: sin-
gle iteration; lower row: iterated until convergence)
training source data and the ground truth, the CA requires more iterations to achieve
convergence (typically about four iterations).
A final example of Rosin's method [34] is given where the USF training data had
salt and pepper noise with probability of 0.1. This enables a set of seven rules to
be learnt that are robust to similar noise. Results are shown in figure 5.8 for edge
detection on a noisy version of the MIT image. Application of the Sobel produces a
very noisy edge map (figure 5.8a), while applying a CA with the original rule learnt
from the clean USF data also fares poorly (figure 5.8b). Using the set of rules learnt
from the noisy training data produces a much better result (figure 5.8c), requiring
on average about fifteen iterations of the rules. However, since the rule learnt for
edge detection is restricted to inverting white pixels then the 'salt' (white noise)
is effectively removed, but the 'pepper' (black noise) remains. The solution taken
in Rosin [34] is to also apply the rules to the inverted image so as to remove the
 
Search WWH ::




Custom Search