A Methodology for Enhancing Pattern Classifiers and Mapping Scores to Probabilities

This research allows one to take an existing pattern classifier (software/hardware system) and enhance its performance in two ways: (i) it maps class scores produced by the classifier to probability values, and (ii) it allows the performance of the classifier to be improved with a new training set without having to access the internal structure of the classifier. Moreover, every time the original classifier improves, this approach can further improve the classification performance. The probability values returned are truly scale independent, thus carrying the same meaning across classifiers. This makes the research extremely attractive for classifier combination applications. The methodology described in this paper has been tested. A word recognizer is used to demonstrate the conversion of confidence (distance) scores to probabilities and digit recognizers are used to demonstrate the improvement in recognition accuracy. Regular Paper Index terms: Recognizer, re-estimation methods, fea..