Edge Detection Using Cellular Automata - Cellular Automata in Image Processing and Geometry - page 97

We now show results for a variety of other CA based edge detectors. In figure 5.9
is shown Wongthanavasu and Lursinsap's [45] method. For a single iteration the
results look reasonable. Additional iterations degrade the results.
Figure 5.10 shows results from Popovici and Popovici's [26] method (which con-
verged after three iterations). 3 They require an application specific threshold param-
eter to be specified, and it can be seen to alter the density of the detected edges. The
best results in figure 5.10 might be considered to be with the threshold set to 16,
although even then the results are generally inferior, containing thick edge regions
whilst also retaining many scattered and disconnected single pixel edges.
Figure 5.11 shows results from Diwakar et al. 's [7]. As expected, since the
method is based on global thresholding, it has missed many edges and found many
spurious ones. If more than a single iteration of the CA is run, the results degrade
even further.
The results of applying a post-processing step of edge linking are shown in fig-
ure 5.12. The input image is edge detected using Rosin's method [34], and then
thresholded to create a binary edge map. In addition, isolated edge pixels were re-
moved. The linking method described in section 5.4.1 was applied for five iterations.
It can be seen that many fragmented edges have been successfully linked (as shown
by the red edgels).

5.6 Conclusions

As stated in the introduction, it is difficult for the general reader to gain an under-
standing of the state of the art in cellular automata based edge detection since papers
are dispersed over many conferences and journals. Our brief survey shows that there
exists a relatively large number of relevant papers, although a number of them were
not clearly written, with details missing or occasionally inconsistent. Moreover, a
number of papers are misleading, in that, according to common usage within image
processing, they actually perform (binary image) boundary detection rather than
(intensity image) edge detection.
The papers and the results of the experiments included in this chapter demonstrate
that cellular automata are indeed capable of performing edge detection, i.e. process
a grey level input image to produce an edge magnitude image (either binary or e.g.
256 values) as output. The results were of variable quality, but in order to be able to
confidently evaluate and compare edge detectors a more systematic and quantitative
analysis should be carried out. This has not been done to date.
The experiments revealed that for all the CA only a very few iterations were
necessary to achieve their optimal results (such details were often missing from the
descriptions of the methods in their original publications). Specifically, the meth-
ods of Rosin [34] (when trained on the USF dataset), Wongthanavasu and Lursin-
sap [45] and Diwakar et al. [7] should generally run for only a single iteration.

3

Popovici and Popovici's [26] paper described a von Neumann neighbourhood, but we
found better results (those shown in figure 5.10) to be obtained using a Moore neigh-
bourhood.