毕业生知识库

联机手写文本识别技术在智能移动设备中得到了越来越多的应用。目前主 流的方法虽然能取得较高的识别准确率，但是其需要合成大量文本行训练数据 或者进行数据增强以提升性能，并且因为无法给出字符切分位置，无法保证良好 的可解释性。同时，目前的方法大多针对水平书写文本行设计，难以用于任意书 写方向的文本行识别。基于过切分的方法模拟人类的认知过程，使用一个单字分 类器对文本行中的字符进行切分和识别，具有较高的可解释性；同时只需要对 过切分模块进行改进便可较好地识别任意书写方向文本行。本文以显式切分为 基本路线，使用卷积原型网络（Convolutional prototype Network, CPN）和双向长 短期记忆网络（Bidirectional Long Short-Term Memory Network, BiLSTM）改进过 切分识别方法实现对任意书写方向文本行的识别并达到具有竞争力的识别性能。 具体贡献如下： 1. 为了缓解联机手写中文文本识别器设计对大量文本行训练数据的需求，提 出了一种利用卷积原型网络的基于过切分的联机手写中文文本行识别方法。为 了进一步提升模型对非字符的拒绝能力，本文使用了字符级的优化目标函数进 行弱监督学习。此外，为了更好地利用字符样本进行训练和识别，本文提出了一 种新的基于文本行几何信息的字符样本归一化方法。与基线模型相比，该模型 在 CASIA-OLHWDB 和 ICDAR2013-Online 数据集上都显著提高了文本行识别性 能；与主流方法相比，识别结果也具有竞争力。证明了本文提出方法的有效性。 2. 为了实现对任意书写方向文本行的识别，本文将任意书写方向的文本行 分为文本行整体旋转和书写方向旋转而文字保持直立两种情况，在预处理阶段 加入一个文本行书写方向检测模块实现对这两种情况的分类。前一种情况经过 旋转矫正后直接进行识别，针对后一种情况，设计了一种基于双向长短期记忆网 络笔画分类模型的改进过切分算法；对任意书写方向文本行进行识别。实验结 果表明本文提出的改进方法实现了过切分框架下对任意书写方向文本行的识别， 并且进一步提升了模型对水平文本行的识别性能。

The technology of online handwritten text recognition has gained increasing application in intelligent mobile devices. Although current mainstream methods can achieve high recognition accuracies, they require synthesizing of large-scale training text data or data augmentation to improve the performance. Additionally, they cannot provide character segmentation positions, resulting in limited interpretability. Furthermore, most existing methods are designed for horizontally written texts and are not suitable for text recognition in arbitrary writing directions. On the other hand, the over-segmentation approach simulates the human cognition, performing character segmentation and recognition using a character classiffer, and can apply to the recognition of arbitrary writing directions. Along the approach of explicit segmentation, this thesis proposed an improved online handwritten Chinese text recognition method using Convolutional Prototype Network (CPN) and Bidirectional Long Short-Term Memory Network (BiLSTM) to achieve competitive recognition performance. The speciffc contributions are as follows. 1. To alleviate the demand for large-scale training text data in online handwritten Chinese text recognition system design, an over-segmentation-based method utilizing a CPN is proposed. To enhance the model’s ability to reject non-character patterns, we design a character-level optimization objective for weakly supervised learning. Additionally, to better utilize character samples for training and recognition, a new geometric information-based character sample normalization method is introduced. Compared to the baseline model, this method signiffcantly improves text recognition performance on the CASIA-OLHWDB and ICDAR2013-Online datasets. Furthermore, the recognition results are competitive compared to state-of-the-art methods. 2. To enable recognition of text lines of arbitrary writing directions, we divide handwritten texts into two categories: texts with overall rotation and texts with writing direction rotation while keeping the text upright. We design a text writing direction detection module to classify these two cases in the pre-processing stage. For the ffrst case, after rotation correction, the text line can be processed using an ordinary text recognizer. For the second case, an improved over-segmentation algorithm is designed based on a BiLSTM stroke classiffcation model to classify strokes of arbitrary writing directions. Experimental results demonstrate that the proposed method achieves high accuracies in character segmentation and recognition of texts of arbitrary writing directions, and the improved over-segmentation method is beneffcial to the recognition performance of horizontal texts.