TABLE OF CONTENTS In this section we will describe how optics can best be used to perform the last major step in the pattern recognition process: classification. In the classical approach to pattern recognition outlined earlier (Fig. 12.1) there are three major steps: sensing, feature measurement/extraction, and classification. Typically, hybrid systems have been conceived where optics is used to reduce the space-bandwidth product of the sensed image to an optimal or suboptimal set of features. Subsequently, the reduced feature set is sensed by a detector (or detector array), thus converting it to an analog electronic signal, which can be subsequently digitized. Further computation for classification can therefore be accomplished by a particular algorithm using a serial (or parallel) digital processor.
The process of developing a classifier is usually an iterative design procedure involving training and testing of the classification algorithm. Statistical measures of effectiveness, such as minimum mean-square error or maximum likelihood, are established by an analysis of the ensemble of feature vectors for a particular classification experiment. The classifier algorithm is often evaluated in terms of the probability of correct classification (as well as probability of incorrect classification) and an associated confidence level.
One of the concerns in classification is the ability of the classifier to verify the correct classification of an object when the sensed object (test signal) is distorted by noise or other systematic distortions, such as translation, scale, and rotation. This is why we were concerned about these earlier when discussing matched-filter correlators, since they...
