A symbol graph based handwritten math expression recognition

摘要: In online handwritten math expression recognition, one-pass dynamic programming can produce high-quality symbol graphs in addition to best sequence hypotheses, especially after discriminative training and trigram graph rescoring. Impact of on whole however, has not been referred yet, since the interface structure analysis module does work well with basis typical tree search. this paper, we propose a method convert segment make search efficient effective, i.e., segmentations without pruning becomes possible. With rescoring, overall recognition accuracy improved by 10% relative comparison baseline.