基于改进卷积神经网络的手写汉字识别研究  

Research on Handwritten Chinese Character Recognition Based on Improved Convolutional Neural Network

作　　者：王建华[1] 张雅祺 肖博怀 李本建[1] WANG Jian-hua;ZHANG Ya-qi;XIAO Bo-huai;LI Ben-jian(College of Arts,Guilin University of Technology,Guilin 541006,China;School of Information Science and Engineering,Guilin University of Technology,Guilin 541006,China)

机构地区：[1]桂林理工大学艺术学院,桂林541006 [2]桂林理工大学信息技术与工程学院,桂林541006

出　　处：《印刷与数字媒体技术研究》2023年第1期45-56,共12页Printing and Digital Media Technology Study

摘　　要：手写汉字识别是智能信息化的重要研究领域之一,在手写输入、文件录入等相关领域至关重要。然而,由于汉字的种类繁多、结构复杂等问题,通常的深度学习方法对汉字的特征提取能力差、计算量大,汉字识别的准确度无法达到令人满意的程度。为解决上述问题,本研究提出了一种改进的深层卷积神经网络模型,在模型的结构上添加了批标准化层,在损失函数上加入了Dropout和正则化方法,在训练过程中加入了RMSprop优化器。为了证实提出的模型的有效性,在CASIA-HWDB1.1和IAHCC-UCAS2016的数据集上进行实验。对提出的模型进行了整体实验,与其他深度学习模型进行了分组对比实验,对批标准化层和RMSprop优化器进行了消融实验,全方面验证了本研究模型的高准确率和高运行速度。Handwritten Chinese character recognition is one of the important research areas of intelligent information technology,which is crucial in handwritten input,document entry and other related fields.However,due to the wide variety and complex structure of Chinese characters,the usual deep learning methods have poor feature extraction ability and large computational effort for Chinese characters,and the accuracy of Chinese character recognition cannot reach a satisfactory level.To solve the above problems,in this study,an improved deep convolutional neural network model was proposed by adding a batch normalization layer to the structure of the model,adding Dropout and regularization methods to the loss function,and adding an RMSprop optimizer to the training process.To confirm the effectiveness of the proposed model,the experiments were conducted on the dataset of CASIA-HWDB1.1 and IAHCC-UCAS2016.The overall experiments of the proposed model were conducted,grouped with other deep learning models,and the ablation experiments of the batch normalization layer and RMSprop optimizer were conducted to verify the high accuracy and high running speed of the proposed model from an all-round perspective.

关 键 词：手写汉字识别 卷积神经网络 深度学习 

分 类 号：TS801.8[轻工技术与工程] TP391.1[自动化与计算机技术—计算机应用技术]