基于深度学习的端到端印刷体数学公式识别
首发时间:2019-12-26
摘要:印刷体数学公式的识别对于在线教育和防止学术不端都有重要的现实意义,目前大多研究者使用的是传统的分步骤方法,实现困难且泛用性不强。本文提出一种基于深度学习的端到端的印刷体数学公式识别方法,使用卷积神经网络(convolutional neural network,CNN)提取图像特征,并使用编码器-解码器结构的循环神经网络(recurrent neural network,RNN),将特征翻译为LaTeX 文本。使用识别前图片和识别的公式结果生成的图片的距离作为评价标准,在基于2003年KDD Cup数据集的提取的公式数据上取得了良好的效果。
关键词: 人工智能 深度学习 卷积神经网络 公式识别 光学字符识别
For information in English, please click here
The End-to-End Recognition of Printed Mathmatical Formulas Based on Deep Learning
Abstract:The recgonition of printed mathematical formulas has important practical significance for both online education and prevention of academic misconduct. At present, most researchers use traditional step-by-step methods, which are difficult to implement and are rarely universal. This article proposes an end-to-end method for recognition of printed mathematical formulas based on deep learning. It uses a convolutional neural network (CNN) to extract image features and uses a recurrent neural network (RNN) with an encoder-decoder structure to translate features into LaTeX text. Using the distance between the picture before recognition and the picture generated by the formula result as the evaluation criterion, good results have been obtained on the formula data extracted based on the 2003 KDD Cup dataset.
Keywords: artificial intelligence deep learning convolutional neural network formula recognition OCR
基金:
引用
No.****
同行评议
勘误表
基于深度学习的端到端印刷体数学公式识别
评论
全部评论0/1000