13.1 Pattern Recognition

159

13.1

Pattern Recognition

Ultimately, pattern is a psychological concept: A set of objects fulfilling conditions

of unity and integrity, according to which groups of objects with some common

feature(s) are denoted and perceived (i.e., distinguished from other objects in their

environment) by a human being. Pattern is therefore synonymous with class, category,

or set. The remark that “a pattern is equivalent to a set of rules ¹ for recognizing it”

is attributed to Oliver Selfridge. Recognition is the process whereby an unknown

object is attributed to a certain pattern (and hence requires the existence of more than

one pattern). The attribution level involves comparison of the unknown with known

objects (prototypes). Features can be qualitative or quantitative (measurable); the

latter are required for automated pattern recognition. The ability to select and rank

features is one of the most complex and important processes of the human intellect,

and it is not surprising that it is perhaps the greatest challenge facing completely

automated computer-based systems. At present, features are typically selected by a

human.

The basic steps of pattern recognition are as follows:

1. Choice of the initial feature set. The number of features determines the dimen-

sionality of feature space.

2. Measurement of the chosen features of a prototype.

3. Preparation (elimination of excess information—noise), ² resulting in a somewhat

standardized description (a prototype), which is then used to construct the training

set.

4. Construction of the decision-making rule.

5. Comparison of any (typically prepared) unknown object with a prototype; with

the help of a quantitative resemblance measure, a decision is made whether the

unknown object belongs to the pattern.

Pattern recognition is thus seen to be a supervised (i.e., undertaken with a teacher)

learning process. Learning implies that the decision-making rule is modified by

experience. The process of pattern recognition is typically computationally heavy;

thus, in this field there is a strong motivation for finding algorithms that are very

efficient.

The discernment of clumps or clusters of objects according to the features chosen

to represent them transcends the recognition of patterns in the sense of noting the

similarity of a known object to an unknown object. Where data are simply analysed

and clusters are found, this is pattern discovery and is dealt with in the next section.

1 That is, an algorithm.

2 For example, imagine a typical time-varying signal such as the output of a microphone. This can

be converted to a square wave of uniform amplitude and varying period.