Tree Transformation and Neural Network Based Hand-written Formula Recognizer

Yuxiang Guo, Yuming Han, Haotong Cao, Kailei Zhu, Jianbo Du

Abstract: Automatic recognition of mathematical expressions is one of the key techniques to improve the access to large amount of document collections and retrieval on the design of fifth generation cellular systems (5G) and future networks. A formula recognizer mainly contains by three parts, such as components separation, components recognition and the connection construction between components, etc. It is common to employ a neural network as the core of the whole system, and the performance can be further improved by combining with tree transformation. In this paper, we considers a novel tree transformation based neural network to analyze the structure of formulas. Then, we propose an error checking method after symbol recognition so that our system is able to distinguish the upper-case and lower-case of a letter. For imbalanced database, we propose to use sampling and thus to obtain an uniform distribution database. Numerical results show that our proposed method performs well in delay reduction and accuracy improvement compared with other existing schemes.