Examination of metrics and assumptions used in correlation filter design

This paper examines some of the metrics that are commonly used to design correlation filter's for optical pattern recognition, including: the Fisher ratio, the signal-to-noise ratio, the equal correlation peak constraint, and normalized correlation. Attention is given to the underlying assumptions that are required to move the Bayesian decision theory to a particular metric or design principle. Since a Bayes classifier is statistically optimum, this provides a means for assessing the merit of a particular approach. Although we only examine a few metrics in this paper, the approach is general and should be useful for assessing the merit and applicability of any of the numerous filter designs that have been proposed in the optical pattern recognition community.