基于卷积循环神经网络的手写汉字文本识别  

Handwritten Chinese Character Text Recognition Based on Convolutional Recurrent Neural Network

作　　者：胡瑞朋 何春燕 张伟明[1,2] 赵立新 李明博 HU Rui-peng;HE Chun-yan;ZHANG Wei-ming;ZHAO Li-xin;LI Ming-bo(Institute of Machinery and Equipment Engineering,Hebei University of Engineering,Handan 056038,China;Key Laboratory of Intelligent Industrial Equipment Technology of Hebei Province,Hebei University of Engineering,Handan 056038,China;Ji Zhi Kang(Beijing)Technology Co.,Ltd.,Beijing 102600,China)

机构地区：[1]河北工程大学机械与装备工程学院,邯郸056038 [2]河北工程大学河北省智能工业装备技术重点实验室,邯郸056038 [3]集智康(北京)科技有限公司,北京102600

出　　处：《科学技术与工程》2025年第4期1547-1554,共8页Science Technology and Engineering

基　　金：河北省教育厅科学研究项目(CXY2024046)。

摘　　要：为了解决卷积循环神经网络(convolutional recurrent neural networks, CRNN)手写汉字文本识别网络模型的训练参数大、文本识别率低等问题,提出一种基于注意力双向长短期记忆网络(based on attention bi-directional long short-term memory network, AT-BLSTM)和知识蒸馏(knowledge distillation, KD)技术的手写汉字识别方法。通过对AT-BLSTM网络的输入向量特征赋予不同的权重,使模型训练数据集更加高效、准确;通过KD技术将一个高性能的大模型获取的知识传输到一个小模型中,在确保模型准确性的同时,减少训练参数和内存占比,得到一个性能更优的轻量级训练模型。该方法通过多组实验对比,汉字识别准确率提高了6.7%,训练参数减少15.94 M。该网络模型识别准确率达到97.9%,汉字识别效果更好。In order to solve the problems of large training parameters and low text recognition rate of convolutional recurrent neural networks(CRNN)handwritten Chinese character recognition network model,a novel method for handwritten Chinese character recognition based on attention bi-directional long short-term memory network(AT-BLSTM)and knowledge distillation(KD)technology was proposed.By assigning different weights to the input vector features of AT-BLSTM network,the model training data set was more efficient and accurate.Through KD technology,the knowledge acquired from a large high-performance model was transferred to a small model,which ensured the accuracy of the model,reduced the training parameters and internal storage ratio,and obtained a lightweight training model with better performance.Through the comparison of multiple groups of experiments,the accuracy of Chinese character recognition is increased by 6.7%,and the training parameters are reduced by 15.94 M.The recognition accuracy of this network model reaches 97.9%,and the recognition effect of Chinese characters is better.

关 键 词：卷积循环神经网络(CRNN) 手写汉字文本识别 注意力机制 知识蒸馏(KD) 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]