Edge Detection Using Cellular Automata - Cellular Automata in Image Processing and Geometry - page 83

detection continues to be an active area. For instance, more recent approaches try to
combine changes in brightness, colour, and texture [18], and use machine learning
to determine how these multiple cues are combined.
During the last decade or so there has been considerable interest in using cellular
automata to perform edge detection. They have several potential benefits, such as:
• Efficiency - Cellular automata are both inherently parallel and computationally
simple, which means that they can implemented very efficiently in hardware.
• Global properties - Since multiple iterations of cellular automata rules can lead
to emergent global behaviour it is feasible that a cellular automaton approach
could provide benefits regarding the tendency of existing methods to produce
fragmented output where there is locally inconsistent data.
• Application specificity - Rules can be selected (e.g. learnt from training data) to
operate better than general purpose edge detectors for demanding applications,
e.g. noisy modalities such as ultrasound and SAR.
There are many applications of CA based edge detection. For instance, in med-
ical image processing some examples are: detection of tumours [7], identification
of the pectoral muscle in mammograms [43] and diagnosis of lung cancer [30].
Other applications include: analysis of antibiotic images [19], detection of fabric
defects [42], detection of grain boundaries in rocks [11], horizon tracking [9] and
content based image retrieval (see chapter 8).
However, it is difficult for the general reader to gain an understanding of the
state of the art in cellular automata based edge detection as the relevant papers are
dispersed over many conferences and journals, which are often not specific to either
cellular automata or image processing/computer vision. This chapter aims to collect
together descriptions of these methods, and to provide a critical assessment of their
merits.
Another obstacle in the path of the reader is the lack of performance evaluation. In
the general area of computer vision there has been considerable work on comparison
of edge detectors, using various methodologies, e.g. human assessment, comparison
to (human) ground truth [12, 13, 40]. However, the majority of papers on developing
cellular automata methods for edge detection simply show a few examples of their
results alongside the Sobel or Canny outputs. Ideally this should be replaced by
quantitative evaluation scores. However, this process is generally complicated by
the need to process the raw outputs of the edge detectors before comparison, and so
would depend on parameters for thresholding, thinning, etc.

5.2 Boundary Detection

The classical or the most popular cellular automata (CA) are binary, so it is natural
to use CA for binary image processing. Edge detection of binary images can be
considered as finding the boundaries of objects or regions in an image, and thus it is
also called boundary detection.
When CA are used for image edge detection, the image to be detected is usu-
ally considered as a cellular automaton, each pixel of which is taken as a cell that