摘要
作为世界六大古文字之一的古彝文记录了几千年来人类的发展历史。通过对古彝文的识别能够将这些珍贵文献资料转换为电子文档,便于保存和传播。由于历史发展、区域限制等原因,针对古彝文识别的研究鲜有成果。本文将当前新颖的深度学习技术应用于古老的文字识别。在四层卷积神经网络(convolutional neural network, CNN)基础上扩展出5个模型,然后利用Alpha-Beta散度作为惩罚项,对5个模型的输出神经元重新进行自编码,接着用2个全连接层完成特征压缩,最后在softmax层对古彝文字符特征进行重新评分,得到其概率分布,选择对应的最高概率作为识别的字符。实验表明,相对于传统CNN模型,本文方法对古彝文手写体的识别精度更高。
Ancient Yi is one of six kinds of ancient word in the world and it recorded human development history in thousands of years.The recognition technology of Ancient Yi character will enable us to transform lots of precious Ancient Yi literature into electronic documents which is convenient for storage and spread.Due to imbalance of historical development and territorial limitation,the research on recognition of ancient Yi is rare.In this article,we apply novel deep learning to ancient character recognition.Our framework consists of five models which are extended based on four-layer convolutional neural network (CNN).Then we take Alpha-Beta divergence as penalty term to implement the coding for output neurons of five models.Next,two fully connected layers finish characteristics compression.Finally,we use softmax layer to reevaluate characteristics of ancient Yi character and get the probability distribution.The character owning the highest probability is identified as the target character.Experiments show that our method has higher precision compared with the traditional CNN model for handwriting recognition of the ancient Yi.
作者
陈善雄
王小龙
韩旭
刘云
王明贵
CHEN Shanxiong;WANG Xiaolong;HAN Xu;LIU Yun;WANG Minggui(Southwest University,Chongqing 400715,China;Department of Information Engineering,Guizhou University of Engineering Science,Bijie 551700,Guizhou Province,China)
出处
《浙江大学学报（理学版）》
CAS
CSCD
北大核心
2019年第3期261-269,共9页
Journal of Zhejiang University（Science Edition）
基金
国家自然科学基金资助项目(61603310)
中国博士后基金项目(2015M580765)
重庆市博士后科研项目(Xm2016041)
中央高校基本科研业务费项目(XDJK2018B020
XDJK2018B019)
关键词
古彝文
深度学习
卷积神经网络
散度
Ancient Yi
deep learning
convolutional neural network
divergence
