Abstract

Given a collection of training images (TIs) to be recognized and a set of filters designed to represent the TIs, a recognition system based on the Bayes likelihood ratio test typically requires a calibration data set to provide an estimate of each probability distribution residing in the signal space. In order to make use of all the correlations evaluated, the dimension of the signal space should be set equal to the product of the number of filters (N_f) and the number of components required by the correlation metric (N_m). Responses are thus N-dimensional vectors (where N = N_fN_m), and characterizing the response distributions (one per TI) by an N-dimensional variance analysis allows an unknown sample to be classified by correlating its response with each of the distributions. The calibration data not only provides descriptions of the distributions (used for classification), but also implicitly locates the boundaries between classification regions. Knowledge of the boundaries permits an evaluation of system performance by integrating classification errors over the entire signal space. A recognition system using the procedure outlined above was constructed and calibrated. Expected performance was calculated and compared to observed performance.

© 1992 Optical Society of America