A new warping technique for normalizing likelihood of multiple classifiers and its effectiveness in combined on-line/off-line japanese character recognition

Abstract

We propose a technique for normalizing likelihood of multiple classifiers prior to their combination. Our technique takes classifier-specific likelihood characteristics into account and maps them to a common, ideal characteristic allowing fair combination under arbitrary combination schemes. For each classifier, a simple warping process aligns the likelihood with the accumulated recognition rate, so that recognition rate becomes a uniformly increasing function of likelihood. For combining normalized likelihood values, we investigate several elementary combination rules, such as sum-rule or max-rule. We achieved a significant performance gain of more than five percent, compared to the best single recognition rate, showing both the effectiveness of our method for classifier combination and the benefit of combining on-line Japanese character recognition with stroke order and stroke number independent off-line recognition. Moreover, our experiments provide additional empirical evidence for the good performance of the sum rule in comparison with other elementary combination rules, as has already been observed by other research groups.