Postprocessing of correlation outputs to improve optical recognition system performance

Abstract

Optical correlators employing magnetooptic spatial light modulators (MOSLMs) can process binary phase-only filter (BPOF) responses at kilohertz frame rates,1 making real-time recognition possible. However, aspect invariant recognition is difficult since simple BPOFs are highly sensitive to geometric distortions. Aspect invariant composite BPOFs have been developed; unfortunately, filters that achieve aspect invariance over wide distortion ranges produce reduced correlation peak values, rendering simple threshold decision rules ineffective. A statistical correlation plane filter (CPF) based on analysis of variance has been shown to improve recognition system performance.2 This paper examines a multiclass quadratic CPF based on a Bayes likelihood ratio test. The quadratic CPF allows the correlation system to be calibrated by repeatedly analyzing correlation peaks for all object–filter combinations. This off-line calibration process yields a vector of expected correlation peak values and an associated covariance matrix for each unique object with each filter in the filter bank. Once gathered, these data are used on-line to calculate the Mahalanobis distance between an observation vector of peak values obtained with an unknown object and each mean vector. Recognition based on the shortest Mahalanobis distance delivers robust performance, even when composite BPOFs are used in the optical correlator.