Recognition of Cursive/Blind Kanji Handwriting Utilizing the Contuinous Speech Recognition Approach

在本项目中，连续语音识别（CSR）方法成功地应用于在线手写识别，并取得了巨大的成果。该技术以25个基本笔画和汉字结构词汇代替汉字拼音中的音素和语法，利用先进的汉字拼音技术在解空间中进行高效搜索，能够识别草书和盲手写汉字。该技术通过观察笔运动的速度矢量序列，来估计正在书写的汉字的强度，而不是识别绘制的图形，从而可以识别甚至人类无法阅读的严重变形的汉字。在建立了上述技术的基本结构之后，我们对一些研究项目进行了研究，如基于速度的特征（比较几个特征参数）、笔压作为特征、笔画上下文相关建模（考虑前后笔画引起的笔画变形）、笔画宏观模型（将相邻的笔画组合成一个宏观笔画）、对写作者的适应（使hmm适应写作者的特征）、盲写识别（看不见的书写），重叠书写识别（同一区域的多个字符），识别速度加快（改进的搜索算法），数字和平假名的弯曲笔画，以及草书，盲写和盲人书写的集合。利用研究成果，我们的“盲人手写通讯”的研发提案得到了石川县的批准，并获得了1亿日元的支持。这两个项目合作良好，收集了超过100万的手写字符数据，包括草书、盲人和视障字符。

In this project, continuous speech recognition (CSR) approach was successfully applied to online handwriting recognition and produced enormous results. Replacing phonemes and grammar in CSR with 25 basic strokes and Kanji structure lexicon and utilizing highly advanced CSR techniques to make an efficient search in the solution space, the newly developed technology is able to recognize cursive and blind hand-written Kanji characters. By observing the velocity vector sequence of the pen movements, this technology estimates the intension of what Kanji character is being written instead of recognizing the drawn figure so that it may recognize heavily deformed characters even human can not read.After establishing the above fundamental structure of the technology, we investigated a number of research items such as velocity-based features (comparing several feature parameters), pen pressure as a feature, stroke context-dependent modeling (considering stroke deformation caused by previous and following strokes), stroke macro model (adjacent strokes are combined to form a macro-stroke), adaptation to writers (adapting HMMs to the writer's characteristics), recognition of blind writing (written without seeing), recognizing overlapped writing (multiple characters on the same area), acceleration of recognition speed (improved search algorithm), curved strokes for numerics and Hirakana, and collection of cursive, blind writing and ones from blind people.Utilizing the research results, our R&D proposal of "Handwritten Communication for the Blind" was approved by Ishikawa Prefecture with a support of 100 million yen. The two projects cooperated well with each other so that more than 1 million handwritten character data were collected including cursive, blind, and sight-impaired characters.