Abstract

This paper presents a segment and recognize approach to recognize Vietnamese online handwritten text, which is inspired from divide and conquer algorithm. First, we propose two segmentation methods to divide a handwritten paragraph into multiple text lines (text line segmentation) and then multiple words (word segmentation). Secondly, an end to end deep neural network model is developed to recognize Vietnamese handwritten words. Our model is derived from the success of the recent deep neural network models for offline handwriting recognition on English, Chinese, and Japanese. Due to the fact that Vietnamese online handwritten patterns commonly consist of many delayed strokes which are caused by diacritic marks, our approach is to render the online patterns to offline images and recognize them by a deep neural network. Although the offline images rendered from the online patterns are not completely same as the real offline images, they are still good enough to recognize. Besides, the proposed line and word segmentation methods have achieved the segmentation accuracy of 96.67% for line segmentation and 89.47% for word segmentation. Using the segmented handwritten words, the connectionist temporal classification loss with combining of convolutional layers and long short term memory layer are employed. The best recognition accuracy is 95.31% for characters and 88.80% for words, which show the promising results and could be improved in future by further research on different neural network structures.

中文翻译：

---

越南语手写文本识别的新方法

摘要

本文提出了一种识别越南语在线手写文本的分段和识别方法，该方法受到分而治之算法的启发。首先，我们提出两种分割方法，将手写段落分割为多个文本行（文本行分割），然后分割为多个单词（分词）。其次，开发了端到端深度神经网络模型来识别越南语手写单词。我们的模型源自最近用于英语、中文和日语离线手写识别的深度神经网络模型的成功。由于越南语在线手写图案通常由许多由变音符号引起的延迟笔画组成，因此我们的方法是将在线图案渲染为离线图像，并通过深度神经网络对其进行识别。尽管从在线图案渲染的离线图像与真实的离线图像并不完全相同，但它们仍然足以识别。此外，所提出的线切分和分词方法实现了线切分96.67%的切分准确率和分词89.47%的切分准确率。使用分割的手写单词，采用结合卷积层和长短期记忆层的联结时间分类损失。字符的最佳识别精度为 95.31%，单词的最佳识别精度为 88.80%，这表明了有希望的结果，并且可以通过对不同神经网络结构的进一步研究来提高未来的性能。