A unified method for augmented incremental recognition of online handwritten Japanese and English text
A unified method for augmented incremental recognition of online handwritten Japanese and English text
复制标题
DOI：
10.1007/s10032-019-00343-y
复制
发表时间：
2019-09
期刊：
影响因子：
--
通讯作者：
C. Nguyen;B. Indurkhya;M. Nakagawa
中科院分区：
文献类型：
--
作者：
C. Nguyen;B. Indurkhya;M. Nakagawa
我们提出了一种统一的方法来增强在线手写日语和英语文本的增量识别，该方法用于书写时的忙或即时识别，以及书写后的懒惰或延迟识别，而不会产生长时间的等待时间。它将用于分割和识别的本地上下文扩展到一系列最近的笔画，分别称为“分割范围”和“识别范围”。识别范围在分割范围内。增强的增量识别在最近的每几个笔划处触发识别，更新分割和识别候选点阵，并增量地在点阵上搜索最佳结果。它还结合了三种技术。第一种是将之前的识别范围中的分割识别候选格重新用于当前的识别范围。第二个是修复未确定的分割点（如果它们在字符/单词模式之间稳定）。第三种是跳过部分候选字符/单词模式的识别。增强增量方法包括利用上述技术在每个新笔划时触发识别的情况。在TUAT-Kondate和IAM在线数据库上进行的实验表明，其在处理时间、等待时间和识别准确率方面均优于批量识别（一次识别文本）和纯增量识别（每次输入笔划识别文本）。
We present a unified method to augmented incremental recognition for online handwritten Japanese and English text, which is used for busy or on-the-fly recognition while writing, and lazy or delayed recognition after writing, without incurring long waiting times. It extends the local context for segmentation and recognition to a range of recent strokes called “segmentation scope” and “recognition scope,” respectively. The recognition scope is inside of the segmentation scope. The augmented incremental recognition triggers recognition at every several recent strokes, updates the segmentation and recognition candidate lattice, and searches over the lattice for the best result incrementally. It also incorporates three techniques. The first is to reuse the segmentation and recognition candidate lattice in the previous recognition scope for the current recognition scope. The second is to fix undecided segmentation points if they are stable between character/word patterns. The third is to skip recognition of partial candidate character/word patterns. The augmented incremental method includes the case of triggering recognition at every new stroke with the above-mentioned techniques. Experiments conducted on TUAT-Kondate and IAM online database show its superiority to batch recognition (recognizing text at one time) and pure incremental recognition (recognizing text at every input stroke) in processing time, waiting time, and recognition accuracy.