Microscope Image Processing

assume a particu lar functional form for the pdf (the Gaussian, for exa mple )

and use the training data only to estimate the parameters. This gives rise to

a parametric classifier. Considerably less training data is then required for

training, and one benefits from the powerful m athemati cs that have been

developed for those case s.

It is often useful to begin a classifier design effort with a classical Bayes

classifier, as described earlier. At the very least this establishes a baseline of

performance against which other types of classifiers can be tested and evaluated.

Further, if the underlying assumptions are met, the Bayes classifier, assuming

Gaussian statistics, may well perform as well as or better than any other

classifier.

Problems arise when