with a practical amount of hardware, however, requires that the pre-
processing and feature extraction mechanism extract the essence of the
patterns to be identified.
In the design of the pattern recognition system, five major
functional divisions must be considered. These phases are illustrated
in the block diagram of Figure 13. The input pattern or pixel is a
vector quantity made up of many components hence the dimensionality
of the input space may be large.
The purpose of signal conditioning, or "preprocessing" is to
provide data preparation and handling, to provide geometric and radio-
metric correction, to provide a convenient input format, to provide
invariance, to provide in many cases a reduction in the dimensionality
of the input data, and most importantlv, to emphasize or enhance
aspects of the input signal which are deemed important.
An almost universal approach to pattern recognition is to extract
properties or features from the original signal, and to perform the
recognition on the feature profile of the input signal. This serves
several functions. First, by reducing the input pattern to its
essential features, the memory required for storing the signatures is
reduced. Secondly, by reducing the input pattern to independent
features, a considerable amount of invariance to exact form is obtained.
Finally, a degree of invariance to noise and background may be achieved.
Most decision mechanisms are based on multivariate discriminate
analysis that partitions measurement space on the basis of the training
set signatures which then allows a decision to be made for an appropriate
classification for each input pixel.
The function of display is attendant with each and every one of
the other functions of a multispectral processing system as one facet
of the man-machine interaction. Interactive controls and commands allow
the intervention of the operator to direct that certain things be done
which would not otherwise be done automatically.
Post-processing of the classification output to provide summaries
or to integrate the results with other data bases of the user is an
important and sometimes unconsidered function. Automatic map-like
displays of the results are the normal approach, but may be of little
use to some users.
The following sections describe the techniques for implementing
these five functions of the multispectral recognition mechanism in
greater detail.
3.2. Processing Techniques
Data processing refers to the procedures, algorithms, and compu-
tations which are applied to the raw sensor output data to transform
it into useful information to the user. Data processing includes
the combination of (1) data formatting, handling, editing, digitizing,
27