[1]林小渝,陈善雄,高未泽,等.基于深度学习的甲骨文偏旁与合体字的识别研究[J].南京师大学报(自然科学版),2021,44(02):104-116.[doi:10.3969/j.issn.1001-4616.2021.02.015]
　Lin Xiaoyu,Chen Shanxiong,Gao Weize,et al.Oracle Radical and Oracle Combined Character RecognitionBased on Deep Learning[J].Journal of Nanjing Normal University(Natural Science Edition),2021,44(02):104-116.[doi:10.3969/j.issn.1001-4616.2021.02.015]
点击复制

基于深度学习的甲骨文偏旁与合体字的识别研究()
分享到：

《南京师大学报（自然科学版）》[ISSN:1001-4616/CN:32-1239/N]

卷:
第44卷
期数:
2021年02期
页码:
104-116
栏目:
·计算机科学与技术·
出版日期:
2021-06-30

文章信息/Info

Title:
Oracle Radical and Oracle Combined Character RecognitionBased on Deep Learning
文章编号:
1001-4616(2021)02-0104-13
作者:
林小渝1; 陈善雄1; 高未泽1; 莫伯峰2; 焦清局3
(1.西南大学计算机与信息科学学院,重庆 400715)(2.首都师范大学甲骨文研究中心,北京 100048)(3.安阳师范学院 计算机与信息工程学院,河南 安阳 455000)
Author(s):
Lin Xiaoyu1; Chen Shanxiong1; Gao Weize1; Mo Bofeng2; Jiao Qingju3
(1.School of Computer and Information Science,Southwest University,Chongqing 400715,China)(2.Oracle Research Center,Capital Normal University,Beijing 100048,China)(3.School of Computer and Information Engineering,Anyang Normal University,Anyang 455000,China)
关键词:
甲骨文识别; 甲骨文偏旁; 迁移学习; 偏旁分析; 非极大值抑制算法
Keywords:
oracle character recognition; ovacle radical; transfer learning; radical analysis; non-maximum suppression
分类号:
TP399
DOI:
10.3969/j.issn.1001-4616.2021.02.015
文献标志码:
A
摘要:
由于甲骨文字形结构多样,异体字较多,其识别一直是甲骨文领域研究的重要问题. 本文首次提出以甲骨文偏旁为识别的基本构件,建立单偏旁和合体结构的甲骨文字符识别方法,提升甲骨文识别的精度. 方法一:根据甲骨文偏旁字形特点,对甲骨文拓片上的合体字进行甲骨文单偏旁最大极值稳定区域的选取,然后,通过改进的BN-LeNet模型识别甲骨文各个偏旁; 方法二:针对甲骨文合体字拓片稀缺的问题,本文提出一种直接对甲骨文合体字进行整体识别的OraNet模型,该模型采用迁移学习的训练策略,对在脱机手写汉字HCL2000数据集预训练的卷积神经网络模型进行参数和结构上的微调,实现迁移得到低层表示和甲骨文合体字集上高层表示的特征融合,以此来提取甲骨文合体字的高级特征. 实验结果表明,BN-LeNet网络对甲骨文单偏旁识别率为96.24%,微调的OraNet模型对甲骨文合体字识别率为 98.58%,从而表明从甲骨文单偏旁的角度进行甲骨文字形识别,可以获得较高的识别精度. 同时本文将甲骨文视为偏旁组合而非整字识别,这使得算法能够识别从未见过的甲骨文新字,即零样本学习,对甲骨文研究有着重要的应用意义.
Abstract:
Due to the complex glyph structure and many variants,the recognition of oracle characters have always been a important problem in relevant field. This paper proposes to use oracle radical as a component to establish a recognition method of oracle radical and oracle combined character,to improve the accuracy of oracle bone script recognition. Method 1:According to the characteristics of the oracle radical,we have selected the maximum extreme stable region(MSER)of the single radical from the oracle bone script,and then put it into the improved BN-LeNet model to recognize; Method 2:In view of the scarcity of oracle combined character,we have proposed an OraNet model that directly recognizes the oracle bone script character. Transfer learning was introduced into model training,so as to extract the high-level features of the oracle bone script,the fine-tuning strategy is implemented to achieve the feature aggregation of low-level representations and high-level representations. The experimental results show that the recognition rate of BN-LeNet network for oracle radical recognition is 96.24%,and the recognition rate of fine-tuned OraNet model for oracle combined character is 98.58%,which shows that considering the oracle bone script recognition from the perspective of oracle radical,higher recognition accuracy can be obtained,At the same time,we treat oracle bone script as a radical combination instead of whole-word,which enables the system to recognize unseen new oracle bone script,i.e.,zero-shot learning. so it has important application significance for Oracle research.

参考文献/References:

[1] 王宇信. 政府推动下的甲骨文研究全面深入发展与弘扬阶段的时代特色[J]. 殷都学刊,2019,40(1):1-6.
[2]国家文物局. 迎接甲骨文发现120周年研讨会在安阳召开[J]. 遗产与保护研究,2019,4(1):80-80.
[3]顾绍通. 基于拓扑配准的甲骨文字形识别方法[J]. 计算机与数字工程,2016,44(10):2001-2006.
[4]周新伦,李锋,华星城,等. 甲骨文计算机识别方法研究[J]. 复旦学报(自然科学版),1996(5):481-486.
[5]李锋,周新伦. 甲骨文自动识别的图论方法[J]. 电子科学学刊,1996(S1):41-47.
[6]栗青生,吴琴霞,王蕾. 基于甲骨文字形动态描述库的甲骨文输入方法[J]. 中文信息学报,2012,26(4):28-33.
[7]张鲁宁,左信,刘建伟. 零样本学习研究进展[J]. 自动化学报,2020,46(1):1-23.
[8]郭心悦,胡沁涵,刘纯平,等. 基于迁移学习和批归一化的菜肴图像识别方法[J]. 计算机应用与软件,2021,38(3):124-133.
[9]ZHANG H G. HCL2000-A large-scale handwritten chinese character database for handwritten character recognition[J]. Icdar barcelona Spain,2009:286-290.
[10]SRIVASTAVA N,HINTON G,KRIZHEVSKY A,et al. Dropout: a simple way to prevent neural networks from overfitting[J]. Journal of machine learning research,2014,15(1):1929-1958.
[11]姚金良,翁璐斌,王小华. 一种基于连通分量的文本区域定位方法[J]. 模式识别与人工智能,2012,25(2):325-331.
[12]冯国进,顾国华,郑瑞红. 基于自适应投影方法的快速车牌定位[J]. 红外与激光工程,2003,32(3):285-287.
[13]陈善雄,韩旭,林小渝,等. 基于MSER和CNN的彝文古籍文献的字符检测方法[J]. 华南理工大学学报(自然科学版),2020,48(6):123-133.
[14]JAMALUDDIN F N,AHMAD S A,NOOR S,et al. Performance of DWT and SWT in muscle fatigue detection[C]//Biomedical engineering & sciences. New York:IEEE,2016.
[15]WANG Z J,ZHAO Y Q,ZHAO C L. Improved MSER pedestrian detection algorithm based on TOF camera[J]. Journal of physics: conference series,2020,1576(1):012013.
[16]GUAN L,CHU J. Natural scene text detection based on SWT,MSER and candidate classification[C]//2017 2nd International Conference on Image,Vision and Computing(ICIVC). IEEE,2017.

备注/Memo

备注/Memo:
收稿日期:2020-10-18.
基金项目:国家社会科学基金项目(19BYY171)、甲骨文信息处理教育部重点实验室项目(OIP2019E009).
通讯作者:陈善雄,博士,副教授,研究方向:模式识别与图像处理. E-mail:304904212@qq.com
更新日期/Last Update: 2021-06-30