Two recognition systems which classify optical correlation data are de
scribed and experimentally compared. Both require a priori estimates o
f multivariate distributions generated from their training images (TIs
). One system uses a quadratic classifier to partition the signal spac
e into regions associated with single TIs and then defines object regi
ons by forming a union of the appropriate TI regions. The other system
uses a composite Bayes' classifier to partition the signal space dire
ctly into regions associated with single objects. Accordingly, it requ
ires object class distributions which it approximates with composite a
lgebraic functions constructed from the TI distribution estimates. Exp
erimental results demonstrate that the composite Bayes' classifier con
sistently outperforms the modified quadratic classifier, albeit margin
ally, when non-TIs are processed.