Content-Based Image Retrieval with Cellular Automata - Cellular Automata in Image Processing and Geometry - page 143

feature vectors must therefore be indexed for fast searches, usually in some tree
structure. The organisation of such an indexing structure is non-trivial [18, 21].
Note that CBIR is particularly useful in medical imaging. Since this is a con-
strained domain, many refinements to general CBIR algorithms occur in this field.
One example is that of unsupervised segmentation of images [11]. Another speciali-
sation is the use of ant colony optimisation techniques for image processing [12, 24].
Ant colony methods are not explicitly covered in this chapter.
From a CA model point of view, the best rule set to choose for a specific CBIR
scenario is of particular interest. In most of the subtasks of CBIR, methods had
been proposed to learn the best rule sets for a given scenario. Again, references to
this topic is given where relevant, but no detailed discussions are given.

8.3 Cellular Automata in Image Processing and Feature

Analysis

A digital image consists of a two-dimensional grid of pixels, and to apply CA to
image processing in general and CBIR in particular, this grid of pixels is mapped
to a two-dimensional CA, where each cell represents one pixel in the image plane.
Furthermore, it is assumed that the individual automata in each cell are identical,
and hence one transition function can be defined for the CA as a whole. For black
and white images, it is sufficient to consider binary CA, where each cell has only the
two values 0 and 1 (representing black and white). For grayscale and colour images,
cells have more values.
The first uses of CA in image processing came about as parallel implementations
to speed up image processing. These were mostly focused on image processing tasks
such as noise reduction, edge detection and template matching [2, 4, 9, 27, 32]. This
section covers the use of CA in image processing, with the goal of establishing a
feature vector for an image.

8.3.1 Noise Reduction

Noise in an image is defined as random variations in the pixels representing an
image, and hence the task of noise reduction is to replace each such random pixel
with a value as close as possible to the pixel value in the original image. Standard
non-CA based techniques for noise reduction include the application of a Gaussian
filter, a mean filter (both linear) and a median filter (nonlinear) [26]. Noise reduction
techniques depend on the type of noise present in the image, as that influences the
algorithm to be applied to remove the noise. The most common types of noise are
impulse (or salt and pepper) noise, uniform noise, and Gaussian noise. In impulse
noise, the noise pixels have either the minimum or the maximum value over all
possible pixel values (for example, in black and white images, a noisy pixel could
have the value 0 or the value 255). On the other hand, uniform noise presents as
noise pixels which may assume any value over the allowed range.