基于深度卷积神经网络的多字体印刷体汉字识别  

Recognition of Printed Chinese Characters Based on Deep Convolutional Neural Network

作　　者：杨艳华[1] YANG Yan-hua(Computer Engineering College,Jimei University,Ximen 361021,China)

机构地区：[1]集美大学计算机工程学院,福建厦门361021

出　　处：《三明学院学报》2021年第3期38-45,共8页Journal of Sanming University

基　　金：福建省教育厅基金(JAT170327);福建省教育厅中青年教师教育科研项目(JZ160139)。

摘　　要：定义由五个卷积层、四个最大池化层、两个应用Dropout的全连接层和Softmax回归层共12层组成的模型,在所有卷积层和全连接层之后进行批标准化处理,加快模型的学习收敛速度。为了获得大规模的印刷体汉字样本对模型进行训练,采用波纹变换、扭曲、平移和旋转等数据增广方式对国标一级字库的3755个汉字图像进行扩充。经过实验,模型在测试数据集上的识别率达到99.864%。This paper proposes a modified model which consists of five convolutional layers, four max-pooling layers,two full connected layers with Dropout and a Softmax classifier layer. To fast learning speed, lots of batch normalization lay ers are used after all convolutional layers and full connected layers. To obtain large scale samples used in model training, this paper proposes a series of data augmentation methods which include a wave distortion, translation, and rotation. Several sets of experiments are carried out and the recognition accuracy rate reached 99.864% on test data set.

关 键 词：印刷体汉字识别 深度学习 卷积神经网络 数据增广 

分 类 号：TP391.43[自动化与计算机技术—计算机应用技术]