Confidence measures in recognizing handwritten mathematical symbols

Abstract

Recent work on computer recognition of handwritten mathematical symbols has reached the state where geometric analysis of isolated characters can correctly identify individual characters about 96% of the time. This paper presents confidence measures for two classification methods applied to the recognition of handwritten mathematical symbols. We show how the distance to the nearest convex hull of nearest neighbors relates to the classification accuracy. For multi-classifiers based on support vector machine ensembles, we show how the outcomes of the binary classifiers can be combined into an overall confidence value. © 2009 Springer-Verlag Berlin Heidelberg.