Development of OCR (optical character recognition) system for scientific documents

本研究开发了一种适用于科学文献的OCR系统，并将其应用于数学期刊的逆向数字化和数学文献的盲文自动转录。目标图像是通过400-600DPI扫描仪从清晰打印的文件中获得的图像。由于目前还没有能够识别数学符号的商用OCR软件，我们开发了自己的OCR引擎。它能识别数学表达式中使用的约450种字符和符号，并能很好地区分字母的斜体和直立字体。对于文本区域的识别，有几种有效的后处理方法利用语言信息来提高识别结果，而在数学表达区域，基于数学符号结构的其他一些不同的后处理方法是有效的。因此，我们为日语和英语文档开发了分离文本区域和数学表达式区域的算法。在数学表达式的结构分析方面，我们提出了一种新的方法，对不同大小的字符和相似字符的识别误差具有鲁棒性。首先通过与代价关系的可能联系构造一个连接字符（符号）的网络，最后利用反映数学表达式全局结构的代价对候选对象进行重新评价，得到数学公式的识别结果作为网络最小代价的生成树。该方法的优点是通过识别树的总代价自动恢复识别的局部误差。我们还开发了用于编辑数学表达式的手写界面，将其作为一个简单的用户界面来纠正数学表达式的识别错误。

In this research, we developed an OCR system adapted to scientific documents, in view of its application to retro-digitization of mathematical journals and automatic Braille transcription of mathematical documents. The target images are those obtained from clearly printed documents by 400-600DPI scanner.Since there is no commercial OCR software which can recognize mathematical symbols, we developed our own OCR engine. It recognizes with about 450 kinds of characters and symbols used in mathematical expressions, and distinguishes well the italic fonts and upright fonts of alphabets.For the recognition of text areas, there are several efficient post-processing methods to improve recognition results using linguistic information, while in mathematical expression areas, some other different methods of post-processing based on the structure of mathematical notations are efficient. Therefore, we developed algorithms to separate text area and mathematical expression areas, for both Japanese and English documents.As for the structure analysis of mathematical expressions, we developed a new method, robust against the recognition errors of characters and similar characters of different sizes. We first construct a network joining characters (symbols) by possible links of relations with cost, Finally, we obtain the result of the recognition of mathematical formulas as the spanning tree of minimum cost of the network, after reevaluating the candidates by using the cost reflecting global structure of the mathematical expressions. The advantage of this method is that local errors of the recognition are recovered automatically by the total cost of the recognition tree.We also developed handwriting interface to edit mathematical expressions to use it as an easy user interface to correct the recognition errors of mathematical expressions.