基于门控卷积和堆叠自注意力的离线手写汉字识别算法研究

Research on Offline Handwritten Chinese Character Recognition Algorithm Based on Gated Convolution and Stacked Self-Attention

摘要: 针对离线手写文本识别(HTR)在自然语言处理领域中的重要性以及其广泛应用于帮助视障用户、人机交互和自动录入等方面的实际需求，本研究提出了一个全新的模型。该模型在门控卷积网络的基础上引入了堆叠自注意力编码器–解码器，用于离线识别手写的汉字文本。由于书写风格的多样性、不同字符之间的视觉相似性、字符重叠以及原始文档中的噪音等挑战，设计准确且灵活的HTR系统具有相当大的难度，特别是当处理较为复杂、包含大量字符的文本时，算法的学习能力显得不足。为了解决这一问题，我们提出的模型包括特征提取层、编码器层和解码器层。其中，特征提取层从输入的手写图像中提取高纬度的不变特征图，而编码器和解码器层则相应地转录出文本。实验结果显示，该模型在HCTD数据集上的字符错误率(CER)为6.72，单词错误率(WER)为11.11；在HCWD数据集上的实验结果CER为6.22和WER为7.17。相对于其他研究者的模型，本文设计的模型在手写汉字识别率上提升了11%。

Abstract: In light of the significance of offline handwritten text recognition (HTR) in the field of natural language processing and its wide-ranging applications in meeting the practical needs of assisting visually impaired users, enabling human-computer interaction, and facilitating automated data entry, this study proposes a novel model. The model integrates the stacked self-attention encoder-decoder on the basis of gated convolution networks for recognizing offline handwritten Chinese characters. Given the challenges posed by diverse writing styles, visual similarities among different characters, character overlap, and noise in original documents, designing an accurate and flexible HTR system is notably difficult, especially when dealing with complex text containing a large number of characters, where algorithms often demonstrate limited learning capabilities. To address this issue, our proposed model comprises feature extraction, encoder, and decoder layers. The feature extraction layer extracts high-dimensional invariant feature maps from the input handwritten images, while the encoder and decoder layers transcribe the text accordingly. Experimental results demonstrate that the model achieves a character error rate (CER) of 6.72 and a word error rate (WER) of 11.11 on the HCTD dataset; and on the HCWD dataset, the CER is 6.22 and the WER is 7.17. Compared to models developed by other researchers, our designed model shows an 11% improvement in handwritten Chinese character recognition accuracy.