Transformer-based method for mathematical expression recognition in document images

Abstract

The recognition of mathematical expressions (MEs) has attracted many researches in recent years. The results of the recognition of MEs allow to develop many useful applications such as mathematical retrieval systems or editing scientific documents. Compared to traditional methods, the encoder-decoder architecture has achieved many advances in the recognition of MEs. The paper presents strategies to improve the recognition accuracy of the recognition of MEs. Firstly, the advanced neural networks that are the DenseNet and the Transformer are employed in the encoder-decoder. Then, the strategy of generation of training dataset of the encoder-decoder model allow to gain better results in the recognition of MEs. The experiments are carried out on the public datasets (Marmot) and the performance comparison with a deep-learning method has shown the significant improvement of the proposed method.