Real-time structural analysis of handwritten mathematical formulas by probabilistic context-free geometry modelling

Robust structural analysis techniques are crucial for the recognition of handwritten mathematical formulas. Therefore we integrated a novel probabilistic structural classification module into a single-stage semantic decoder recursively taking relative symbol (or symbol group) positions and sizes into account. On the basis of Gaussian, statistically independent distributions for the resulting features we obtained a reliable structural assessment measure which is set against the corresponding symbol classification scores in a top-down chart parsing scheme. The final recognition system performs at a writer specific accuracy of 95.0%. Real-time processing is enabled due to an incremental breadth-first search strategy