Research on High Performance Mathematical Expression Recognition Method for Digitization of Scientific Documents

本研究在原有数学公式识别系统的基础上，开发了一个新的数学公式识别系统，并对数学期刊进行了大量的实验。我们改进的对象有以下几个；(1)收集大量数学表达式中使用的字符和符号。(2)改进处理接触和分离字符或符号的算法。(3)识别各种矩阵和条件表达式。(4)改进脚本表达式识别。(5)发展数学表达式识别的定量性能评估方法。(6)建立一个宏大的真值数据库。对于本课题(1)，从我们与德国埃森大学合作的期刊《档案数学卷60-65》中收集各种字符和符号。对于(2)，做了一些改进，但这是我们未来的工作之一。对于(3)，包括其他矩阵或缩写符号作为元素的矩阵可以很好地识别。对于(4)，由于构造了新的小字体词典，提高了对长文字表达的识别率。对于(5)，每个子表达式的识别结果以MathML格式给出，并自动与本课题(6)开发的宏真值数据库进行比较。该方法使我们能够高效地开发数学表达式识别系统。结果，我们确认了我们的识别系统可以用于实际应用。

In this research, a new mathematical expression recognition system based on our former system has been developed, and intensive experimentations were curried out for mathematical journals. The subjects of our improvement are as follows;( 1 ) Collection of a large numbers of characters and symbols used in mathematical expressions.( 2 ) Improvement of a algorithm which can deal with touching and separated characters or symbols.( 3 ) Recognition of wide variety of Matrices and conditional expressions.( 4 ) Improvement of script expression recognition.( 5 ) Development of a quantitative performance evaluation method for mathematical expression recognition.( 6 ) Construction of a grand truth database.For the subject ( 1 ), various characters and symbols were collected from the journals "Archiv der Mathematik vol. 60-65" which is the documents we are working on the joint project with Essen University, Germany. For ( 2 ), some improvements were made, but this is one of our future works. For ( 3 ), matrices including other matrices or abbreviation symbols as elements can be recognized fairlywell. For ( 4 ), the recognition rate for long script expressions was improved due to construction of anewdictionaryof small sized fonts. For ( 5 ), recognition results for each sub-expressions are presented in mathML formats, and compared automaticallywith the grand truth database developed in the subject ( 6 ). This method enables us to develop our mathematical expression recognition system efficiently. As a result, we confirmed that our recognition system can be used for real applications.