The answer to the three problems described in Subsection 6.6 clearly is, to partition the pattern space in a manner which requires the determination of fairly few constants. The few constants usually can be evaluated from the available representative patterns (which in most cases are limited in number) with a reasonable effort in time and funding, and the values of the few constants can be stored in memories of realistic size. This section about categorizer design is divided into four parts, one about methods for explicit partitioning of the pattern space, Subsection 7.2, and one about methods for implicit partitioning of the pattern space, Subsection 7.3. In both parts some methods the so called non-parametric methods, will be mentioned where the designer makes no assumption about the distribution in pattern space of the members of the Nc classes. With the remaining methods, the parametric methods, the designer in some manner makes use of known or assumed distributions. In Subsection 7.4 several methods are described for categorization of members from more than two classes, N c > 2. As a preliminary it is in Subsection 7.1 considered how the N c multivariate density functions may be estimated.
KeywordsDiscriminant Function Pattern Point Pattern Space Separation Surface Representative Pattern
Unable to display preview. Download preview PDF.