[1]李晓亮,李光亚,孟志琳. 基于Swin-UNet的破损碑刻文字识别方法[J].机械与电子,2026,44(01):28-34.
　LI Xiaoliang,LI Guangya,MENG Zhilin. A Character Recognition Method for Damaged Inscriptions Based on Swin-UNet[J].Machinery & Electronics,2026,44(01):28-34.
点击复制

 基于Swin-UNet的破损碑刻文字识别方法()
分享到：

《机械与电子》[ISSN:1001-2257/CN:52-1052/TH]

卷:
44
期数:
2026年01期
页码:
28-34
栏目:
智能检测
出版日期:
2026-01-27

文章信息/Info

Title:
 A Character Recognition Method for Damaged Inscriptions Based on Swin-UNet
文章编号:
1001-2257(2026)01-0028-07
作者:
 李晓亮; 李光亚; 孟志琳
 (中北大学信息与通信工程学院,山西 太原 030051)
Author(s):
 LI Xiaoliang; LI Guangya; MENG Zhilin
 (School of Information and Communication Engineering,North University of China,Taiyuan 030051,China)
关键词:
碑刻文字; 文字识别; Swin Transformer; U Net; 语义分割
Keywords:
 ; stone inscription characters; character recognition; Swin Transformer; U Net; semantic segmentation
分类号:
TP18;K877.
文献标志码:
A
摘要:
 提出了基于Swin-UNet的破损碑刻文字识别方法。为了能够获取准确的碑刻文字信息,采用Swin Transformer结构代替U Net结构在分割任务中的下采样和上采样过程,并在其中添加了优化融合特征信息的注意力模块CBAM 与SENet模块,同时使用带权重的交叉熵损失函数对损失函数进行优化。自然场景下的碑刻文字往往会受到各种各样的损害,故之后在数据集的基础上建立文字的语义分割数据库,同时设计算法对缺损的碑刻文字基于数据库进行识别。实验表明,在真实碑刻图片中,文字缺失2个笔画以内,识别正确率为32.60%,识别结果前5个文字中有正确的汉字视为识别正确的概率为64.20%,识别结果前10个文字中有正确的汉字视为识别正确的概率为77.20%。所提方法相较于其他的语义分割模型对笔画的分割更为准确,效果更好。
Abstract:
This paper presents a method for recognizing damaged stone inscription characters based on Swin-UNet.To accurately extract textual information from stone inscriptions,the Swin transformer architecture
is employed to replace the down sampling and up sampling processes of the original U-Net structure in the segmentation task.The CBAM (Convoluted Basin Aggregation Module) and SENet modules are integrated to optimize the feature fusion.The loss function is also refined using a weighted cross entropy loss.Stone inscription characters in natural environments often suffer from various forms of degradation.Consequently,a semantic segmentation database for these characters is constructed based on realworld data sets,and an algorithm is designed to identify damaged characters by leveraging this database.Experimental results demonstrate that on real stone inscription images,the recognition accuracy reaches 32.60% for missing up to two strokes,64.20% when identifying the first five correct characters,and 77.20% for the first ten characters.Compared with other semantic segmentation models,the proposed method achieves more accurate stroke level segmentation and yields superior performance.

参考文献/References:

 [1] 胡娜.数字化如何助力曲阜汉魏碑刻的历史传承[J].文物鉴定与鉴赏,2025,291(3):166- 169.
[2] 王宁.数字化时代的碑刻与碑刻学研究[J].陕西师范大学学报(哲学社会科学版),2017,46(2):119-121.
[3] 李金霞.梧州碑刻现状及保护对策研究[J].桂林博物馆文集,2024,11:248-260.
[4] 邓杰荣,梁森,曹昕妍,等.基于深度学习的汉字识别方法研究综述[J].微纳电子与智能制造,2020,2(3):73-81.
[5] 孙华,张航.汉字识别方法综述[J].计算机工程,2010,36(20):194-197.
[6] ZHANG Y,LUO F F,LIANG W C,et al.EXPRESS:stroke features in the Chinese character recognition[J].Quarterly journal of experimental psychology(2006),2025:17470218251357441.
[7] CRIMINISI A,PEREZ P,TOYAMA K,et al.Region filling and object removal by exemplar based image inpainting[J].IEEE Transactions on image processing,2004,13(9):1200-1212.
[8] HAYS J,EFROS A A.Scene completion using millions of photographs[J].ACM Transactions on graphics,2007,26(3):87-94.
[9] BARNES C,SHECHTMAN E,FINKELSTEIN A,et al.PatchMatch:a randomized correspondence algorithm forstructural image editing[J].ACM Transaction on graphics,2009,28(3):1-11.
[10] KAWAI N,SATO T,YOKOYA N.Diminished reality based on image inpainting considering background geometry[J].IEEE Transactions on visualization and computer graphics,2015,22(3):1236-1247.
[11] 石佳钰.基于生成对抗网络的手写蒙古文字元识别研究[D].呼和浩特:内蒙古师范大学,2022.
[12] 陈善雄,朱世宇,熊海灵,等.一种双判别器GAN 的古彝文字符修复方法[J].自动化学报,2022,48(3):853-864.
[13] 刘栋斌,王慧琴,王可,等.基于内容风格迁移的残缺稀疏文字图像盲修复方法[J].激光与光电子学进展,2022,59(24):106-117.
[14] TANG H,LIU H,XU D,et al.AttentionGAN:unpaired image to image translation using attention guided generative adversarial networks[J].IEEE Transactions on neural networks and learning systems,2023,34(3):1972-1987.
[15] SU B P,LIU X X,GAO W Z,et al.A restoration method using dual generate adversarial networks for Chinese ancient characters[J].Visual informatics,2022,6(1):26-34.
[16] ZHENG W J,SU B P,FENG R Q,et al.EA GAN:restoration of text in ancient Chinese books based on an example attention generative adversarial network[J].Heritage science,2023,11:42.
[17] CAO S X,PAN T W,WANG Y Y,et al.Character restoration of Qin and Han bamboo slips based on improved conditional generative adversarial networks[J].Heritage science,2025,13:278.
[18] 李鸿亮,刘禹良,廖文辉,等.大模型时代的光学文字识别:现状及展望[J].中国图象图形学报,2025,30(6):2023-2050.
[19] 刘栋斌.基于深度学习的古代残缺文字图像修复方法研究[D].西安:西安建筑科技大学,2022.
[20] 孙凯明,郝明,王刚,等.面向历史档案图像的文字修复系统架构设计研究[J].黑龙江科学,2024,15(23):87-89.
[21] SONI K V,SHUKLA V,TANDAN R S,et al.Performance evaluation of efficient and accurate text detection and recognition in natural scenes images using EAST and OCR fusion[J].International journal of advanced computer science and applications (IJACSA),2025,16(1):445-453.
[22] 贺梦兰.基于图网络的残损文字识别[D].西安:西安理工大学,2023.
[23] ASHISH V,NOAM S,NIKI P,et al.Attention is all you need[C]∥ 31st International Conference on Neural Information Processing Systems,2017:6000 6010.

备注/Memo

备注/Memo:
收稿日期:2025-09-04
基金项目:科技部国家重点研发计划(2020YFB2009102)
作者简介:李晓亮 (2000-),男,山西晋中人,硕士研究生,研究方向为数字图像处理与模式识别;李光亚 (1980-),男,山西临汾人,博士,副教授,研究方向为图形图像处理、计算机视觉等,通信作者,E-mail:40827562@qq.com。
更新日期/Last Update: 2026-03-09