何 凯，黄婉蓉，刘 坤，高圣楠
AuthorsHTML:何 凯，黄婉蓉，刘 坤，高圣楠
AuthorsListE:He Kai，Huang Wanrong，Liu Kun，Gao Shengnan
AuthorsHTMLE:He Kai，Huang Wanrong，Liu Kun，Gao Shengnan
Unit:天津大学电气自动化与信息工程学院，天津 300072
Unit_EngLish:School of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China
Abstract_Chinese:手写体中文的自动识别是中文文档数字化的前提和基础，由于中文字符数目繁多、相似性强、字体种类繁多、书写随意、缺乏统一规范等原因，一直是计算机视觉领域中一个具有挑战性的问题．为解决这一难题，提出了一种基于卷积神经网络的手写体中文识别方法．在经典LeNet-5网络模型的基础上进行改进，提出了一种LeNet-Ⅱ模型．利用改进的Inception模块和空洞卷积，设计了一种并行的双路卷积神经网络结构；两路分支可分别提取手写中文图像中不同尺度的特征，获得多个尺度的特征图像；通过对其进行特征融合，可以达到丰富特征图像多样性、提升识别准确率的目的；最后经过全连接层进行分类．利用经典手写体中文数据集进行训练，利用该模型实现了3755类手写体中文字符及相关文本的自动识别．实验结果表明，基于改进LeNet-5模型的手写体中文识别方法，在同一训练数据集上的收敛速度和识别准确率明显优于经典LeNet-5模型，对经典数据集的识别准确率可以达到95.21%，也高于其他传统算法；此外，对4幅手写体中文文本的平均识别准确率达到97.30%，超出了人类表现，取得了理想的实际效果．
Abstract_English:Chinese handwriting recognition is the premise and foundation of the digitization of Chinese documents. It is a challenging problem in the computer vision field for many reasons such as the large number of Chinese characters，strong similarity，wide variety of fonts，free writing forms，and lack of standardization. To solve this problem，a Chinese handwriting recognition approach using a convolutional neural network(CNN)is proposed. An improved LeNet-Ⅱ network model based on the traditional LeNet-5model is also proposed. Using the improved Inception module and dilated convolution，a collateral dual CNN structure is designed. The two branches contribute to the feature extraction of Chinese handwriting images in different scales. After feature fusion，the diversity of feature images is enriched and therefore，the recognition accuracy is improved. Finally，classification is realized using a fully connected layer. Using the proposed network model，3755 handwritten Chinese characters as well as relative Chinese documents and texts are recognized by training the classic Chinese handwriting dataset. The experimental results compared with those obtained using classic methods show that the improved model performs well in terms of convergence speed and recognition accuracy under the same training dataset. The recognition accuracy on the classic dataset is 95.21%，which is higher than those obtained by other traditional algorithms. Moreover，the average recognition accuracy on four handwritten Chinese documents is 97.30%，which is higher than that of human performance.
Keyword_Chinese:手写体中文识别；卷积神经网络；LeNet-5 模型；Inception 模块
Keywords_English:Chinese handwriting recognition；convolutional neural network(CNN)；LeNet-5 model；Inception module

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6501