1.2 Process of Classifier Design
分类器设计流程
Figure 1-6 shows a flow chart of how a classifier is designed. After data
图1-6显示了一个分类器的设计流程图.
is gathered, samples are normalized and registered. Normalization and registration
数据收集好后,样本被规范化和配准.
are very important processes for a successful classifier design. However,
different data requires different normalization and registration, and it is
difficult to discuss these subjects in a generalized way. Therefore, these subjects
are not included in this book.
After normalization and registration, the class separability of the data is
规范化和配准后,每个类的数据被单独处理.
measured. This is done by estimating the Bayes error in the measurement
space. Since it is not appropriate at this stage to assume a mathematical form
for the data structure, the estimation procedure must be nonparametric. If the
Bayes error is larger than the final classifier error we wish to achieve (denoted
by E ~ ) ,th e data does not carry enough classification information to meet the
specification. Selecting features and designing a classifier in the later stages
merely increase the classification error. Therefore, we must go back to data
gathering and seek better measurements.
Only when the estimate of the Bayes error is less than E,,, may we
proceed to the next stage of data structure analysis in which we study the
characteristics of the data. All kinds of data analysis techniques are used here
which include feature extraction, clustering, statistical tests, modeling, and so
on. Note that, each time a feature set is chosen, the Bayes error in the feature
space is estimated and compared with the one in the measurement space. The
difference between them indicates how much classification information is lost
in the feature selection process.
Once the structure of the data is thoroughly understood, the data dictates
which classifier must be adopted. Our choice is normally either a linear, quadratic,
or piecewise classifier, and rarely a nonparametric classifier. Nonparametric
techniques are necessary in off-line analyses to carry out many
important operations such as the estimation of the Bayes error and data structure
analysis. However, they are not so popular for any on-line operation,
because of their complexity.
After a classifier is designed, the classifier must be evaluated by the procedures
discussed in Chapter 5. The resulting error is compared with the
Bayes error in the feature space. The difference between these two errors indicates
how much the error is increased by adopting the classifier. If the difference
is unacceptably high, we must reevaluate the design of the classifier.
At last, the classifier is tested in the field. If the classifier does not
perform as was expected, the data base used for designing the classifier is different
from the test data in the field. Therefore, we must expand the data base
and design a new classifier.