Detecting Mathematical Expressions in Scientific Document Images Using a U-Net Trained on a Diverse Dataset

Abstract

A detection method for mathematical expressions in scientific document images is proposed. Inspired by the promising performance of U-Net, a convolutional network architecture originally proposed for the semantic segmentation of biomedical images, the proposed method uses image conversion by a U-Net framework. The proposed method does not use any information from mathematical and linguistic grammar so that it can be a supplemental bypass in the conventional mathematical optical character recognition (OCR) process pipeline. The evaluation experiments confirmed that (1) the performance of mathematical symbol and expression detection by the proposed method is superior to that of InftyReader, which is state-of-the-art software for mathematical OCR; (2) the coverage of the training dataset to the variation of document style is important; and (3) retraining with small additional training samples will be effective to improve the performance. An additional contribution is the release of a dataset for benchmarking the OCR for scientific documents.

Keywords

• Character recognition
• neural networks
• object detection