Performance prediction for handwritten word recognizers and its application to classifier combination

This paper introduces a performance prediction model for handwritten word recognizers. This model considers the factors involved in word recognition, i.e., the recognizer, input images and lexicons, and presents a quantitative formula to associate performance with these factors. It produces a direct measure of recognition difficulty by the predicted performance which can be utilized to improve the combination of multiple recognizers. We support the accuracy of our model by extensive experiments conducted on five word recognizers and its applications to multiple classifier systems.