XUE Chunhan,JIN Xiaofeng.Few - shot optical characters recognition method of Korean historical document based on transfer learning[J].Journal of Yanbian University,2021,47(04):350-355.
基于迁移学习的少样本朝鲜语古籍文字的识别方法
- Title:
- Few - shot optical characters recognition method of Korean historical document based on transfer learning
- 文章编号:
- 1004-4353(2021)04-0350-06
- Keywords:
- optical character recognition; few - shot learning; data augmentation; generative adversarial network; transfer learning
- 分类号:
- TP391.4
- 文献标志码:
- A
- 摘要:
- 为解决少样本朝鲜语古籍文字识别精度低的问题,提出了一种基于迁移学习的少样本文字识别方法.首先提出了一种结合传统数据增强和条件深度卷积生成对抗网络的数据增强方法,以此扩充朝鲜语古籍文字图像的训练样本数.其次,将富样本集预训练得到的模型迁移到少样本数据集的学习任务中,以此实现少样本的朝鲜语古籍文字识别.实验结果表明,提出的数据增强方法能够满足模型预训练和少样本的学习要求,且VGG16、ResNet18和ResNet50 3种网络模型在测试集上均获得良好的识别性能,其中ResNet50的识别准确率最高(99.72%).因此,该方法可有效解决小样本的朝鲜语古籍文字识别问题,并可为其他语种的小样本文字识别提供参考.
- Abstract:
- To solve issue of low recognition accuracy in Korean historical document characters recognition with lacking samples, a novel method of transfer learning based few - shot optical character recognition is proposed.First, data augmentation fusing traditional and CDCGAN methods is exploited to expand scale of character - segmented image samples.Second, for few - shot training task implementation with homologous - source tactics, transfers the pre - trained model which is obtained by using abundant dataset to accomplish few - shot optical character recognition of Korean historical document. Experimental results show that proposed data augmentation method meets the requirements of pre - training and few - shot learning.VGG16, ResNet18 and ResNet50 outstand performance on test dataset, and ResNet50 achieves the top accuracy at 99.72%.Therefore, the proposed method is aeffective solution to solveisuue of few - shot Korean historical document optical characters recognition, and the method is valuable for same issue in other languages.
参考文献/References:
[1] 杨凡.大数据框架下古籍数字化发展趋势研究[J].图书馆学刊,2017,39(9):74 - 77.
[2] 张亚倩.卷积神经网络研究综述[J].信息通信,2018,191(11):27 - 29.
[3] 柴伟佳,王连明.卷积神经网络的多字体汉字识别[J].中国图像图形学报,2018,23(3):410 - 417.
[4] 陈善雄,王小龙,韩旭,等.一种基于深度学习的古彝文识别方法[J].浙江大学学报(理学版),2019,46(3):4 - 12.
[5] GOODFELLOW I, POUGET - ABADIE J, MIRZA M, et al.Generative adversarial nets[C]//Advances in Neural Information Processing Systems.Cambridge: MIT Press, 2014:2672 - 2680.
[6] 俞彬.基于生成对抗网络的图像类别不平衡问题数据扩充方法[D].广州:华南理工大学,2018.
[7] PAN S J, YANG Q.A survey on transfer learning[J].IEEE Transactions on Knowledge & Data Engineering, 2010,22(10):1345 - 1359.
[8] ZEILER M D, FERGUS R.Visualizing and understanding convolutional networks[C]//European Conference on Computer Vision.Berlin: Springer, 2014:818 - 833.
[9] TANG Y, WU B, PENG L, et al.Semi - supervised transfer learning for convolutional neural network based Chinese character recognition[C]//2017 14th IAPR International Conference on Document Analysis and Recognition(ICDAR).Piscataway: IEEE, 2018:441 - 447.
[10] JANGID M, SRIVASTAVA S.Handwritten devanagari character recognition using layer - wise training of deep convolutional neural networks and adaptive gradient methods[J].Journal of Imaging, 2018,4(2):41.
[11] 刘星辰,金小峰.朝汉混排古籍的文字切分方法[J].计算机工程与应用,2020,56(11):135 - 141.
[12] JIA D, WEI D, SOCHER R, et al.ImageNet: A large - scale hierarchical image database[C]//IEEE Conference on Computer Vision and Pattern Recognition.Miami, FL, USA: IEEE, 2009:248 - 255.
备注/Memo
收稿日期: 2021-04-06
基金项目: 吉林省教育厅“十三五”科学技术项目(JJKH20191126KJ); 延边大学外国语言文学世界一流学科建设项目(18YLPY14)
第一作者: 薛春寒(1996—),女,硕士,研究方向为计算机视觉.
通信作者: 金小峰(1970—),男,硕士,教授,研究方向为语音信息处理、计算机视觉、机器人技术.