Offline handwritten formula recognition is a challenging task due to the variety of handwritten symbols and two-dimensional formula structures. Recently, the deep neural network recognizers based on the encoder-decoder framework have achieved great improvements on this task. However, the unsatisfactory recognition performance for formulas with long $\text{LaTeX}$ strings is one shortcoming of the existing work. Moreover, lacking sufficient training data also limits the capability of these recognizers. In this paper, we design a multimodal dependence attention (MDA) module to help the model learn visual and semantic dependencies among symbols in the same formula to improve the recognition performance of the formulas with long $\text{LaTeX}$ strings. To alleviate overfitting and further improve the recognition performance, we also propose a new dataset, Handwritten Formula Image Dataset (HFID), which contains 25620 handwritten formula images collected from real life. We conduct extensive experiments to demonstrate the effectiveness of our proposed MDA module and HFID dataset and achieve state-of-the-art performances, 63.79% and 65.24% expression accuracy on CROHME 2014 and CROHME 2016, respectively.

Offline handwritten mathematical expression recognition is a challenging optical character recognition (OCR) task due to various ambiguities of handwritten symbols and complicated two-dimensional structures. Recent work in this area usually constructs deeper and deeper neural networks trained with end-to-end approaches to improve the performance. However, the higher the complexity of the network, the more the computing resources and time required. To improve the performance without more computing requirements, we concentrate on the training data and the training strategy in this paper. We propose a data augmentation method which can generate synthetic samples with new LaTeX notations by only using the official training data of CROHME. Moreover, we propose a novel training strategy called Shuffled Multi-Round Training (SMRT) to regularize the model. With the generated data and the shuffled multi-round training strategy, we achieve the state-of-the-art result in expression accuracy, i.e., 59.74% and 61.57% on CROHME 2014 and 2016, respectively, by using attention-based encoder-decoder models for offline handwritten mathematical expression recognition.