基于CRNN改进的中文手写体文本行识别  

Improved Chinese Handwritten Text Line Recognition based on CRNN

作　　者：舒珊珊 郑晓旭 文成玉[1] SHU Shanshan;ZHENG Xiaoxu;WEN Chengyu(College of Communicating Engineering,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学通信工程学院,四川成都610225

出　　处：《成都信息工程大学学报》2023年第4期422-428,共7页Journal of Chengdu University of Information Technology

摘　　要：中文手写体文本行识别可以将纸质书写内容转换为可编辑的电子内容。对于手写体书写随意性大、中文字符种类多,且基于字符分割的方法识别准确率不高这些问题,提出基于卷积循环神经网络改进的端到端的中文手写体识别方法。首先将图片传入基于改进的Inception结构的特征提取网络,该网络首先改进GoogLeNet模型,然后在此基础上又改进添加卷积模块的注意力机制模块和Inception组合结构,改进后的模型能更好地提取图片的有效特征;之后将提取到的图片特征传入循环层,即两层双向长短时记忆网络进行预测;最后将预测序列传入转录层,经过连接时序分类进行转录输出。在CASIA-HWDB2数据集的实验结果表明,该方法能获得95.12%的识别准确率,证明方法的可行性。Chinese handwritten text line recognition converts paper writing into editable electronic content.For the prob-lems of random handwriting,the variety of Chinese characters,and low recognition accuracy of the method based on character segmentation.This paper proposes an improved end-to-end Chinese handwriting recognition method based on Convolutional Recurrent Neural Network(CRNN).First,the picture is passed to the feature extraction network based on the improved Inception structure,the network first improved the GoogLeNet model,and then added the attention mechanism module(CBAM)and the Inception combined structure,after the improvement the model can do better in extracting the effective features of the picture.Then the extracted picture features were passed to the recurrent layer,a two-layer bidirectional long-short-term memory network(BiLSTM),for prediction.Finally,the predicted sequence was passed to the transcription layer,the Connectionist Temporal Classification(CTC),for transcriptional output.Experi-ments use the CASIA-HWDB2 dataset.The results show that the method can obtain a recognition accuracy of 95.12%,which proves the feasibility of the method.

关 键 词：手写体识别 卷积循环神经网络 卷积模块的注意力机制模块 双向长短时记忆网络 连接时序分类 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]