13.1. Theoretical background
Pattern recognition, one of the most important aspects of artificial intelligence, is an appropriate field for the development, validation and comparison of different learning techniques: statistical or structural, supervised or unsupervised, inductive or deductive, etc.
Patterns are general concepts describing an object (mechanical part, obstacle, human face, etc.) or a phenomenon (disease, system operating state, emotion, etc.). They are provided by a sensor or transducer to a recognition system in the form of a data set. The classification is possible only if characteristic information about the observed pattern is confined in this data set, which is then called a pattern signature or fingerprint.
Since the signatures are generally measured, they have a random nature. Thus, the statistical approach, which is described in this chapter, is the most used for pattern recognition. More precisely, we consider the supervised learning framework, which requires a database. This database contains labeled patterns belonging to the M predefined classes ω1 ω2, ..., ωM. These patterns are repeatedly presented to the classifier in order to derive a decision rule optimizing a given criterion.
The general structure of a supervised statistical classification chain is presented in Figure 13.1. This process aims at classifying any unknown pattern using its signature and consists of the following steps:
– feature vector extraction from the recorded ...