LIU Xiaotong,ZHAO Mengling,WANG Guirong,et al.Research on unsupervised clustering method of Korean ancient book character images based on DeepCluster[J].Journal of Yanbian University,2023,(02):183-188.
基于DeepCluster的朝鲜语古籍文字图像的无监督聚类方法研究
- Title:
- Research on unsupervised clustering method of Korean ancient book character images based on DeepCluster
- 关键词:
- 无监督聚类; 朝鲜语古籍; DeepCluster; AlexNet卷积网络; 深度学习; 图像数据集; 文字图像
- Keywords:
- unsupervised clustering; Korean ancient book; DeepCluster; AlexNet convolutional network; deep learning; image dataset; character image
- 分类号:
- TP391.1
- 文献标志码:
- A
- 摘要:
- 为了提高朝鲜语古籍文字图像的标注效率,提出了一种基于DeepCluster的朝鲜语古籍文字图像的无监督聚类方法.首先,基于DeepCluster对AlexNet卷积网络进行简化;然后,采用Sobel滤波器的线性变换消除图像域中的颜色和增加局部图像的对比度;最后,利用数据增强方法强化模型对朝鲜语古籍样本特征的学习能力.在无标注的朝鲜语古籍文字图像数据集上进行实验显示,该方法的准确率和NMI指标比DCN 方法分别提高了15.32个百分点和0.180.由此表明,该方法可有效提高文字图像的标注效率,可应用于朝鲜语
- Abstract:
- An unsupervised clustering method based on DeepCluster for Korean ancient text images was proposed for improve the tagging efficiency of ancient Korean text images.Firstly, the AlexNet convolutional network was simplified based on DeepCluster.Secondly, the linear transformation of Sobel filter was used to eliminate the color in the image domain and increase its local contrast.Finally, data enhancement methods were used to enhance the model for feature learning of Korean ancient text samples.Experiments on unlabeled Korean ancient text images dataset show that the accuracy and NMI metrics of the method improve 15.32 percentage points and 0.180, respectively, compared with the DCN method, thus indicating that the method can effectively improve the efficiency of text image labeling and can be further applied to the construction of Korean ancient character annotation dataset.
参考文献/References:
[1] 苏向东.基于深度学习和知识策略的蒙古文古籍识别研究[D].呼和浩特:内蒙古大学,2016.
[2] 姜丽.基于BIRCH 和改进k中心点算法的古籍汉字图像聚类研究[D].保定:河北大学,2012.
[3] YANG B,FU X,SIDIROPOULOS N D,et al.Towards K-means-friendly spaces:simultaneous deep learning and clustering[C]//International Conference on Machine Learning.Sydney:PMLR,2017:3861-3870.
[4] 王畅,金璟璇,金小峰.聚类与跟踪相结合的人脸数据集生成方法研究[J].延边大学学报(自然科学版),2019,45(3):221-227.
[5] YAN X,MISRA I,GUPTA A,et al.Clusterfit:Improving generalization of visual representations[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Seattle:ICCV,2020:6509-6518.
[6] CARON M,BOJANOWSKI P,JOULIN A,et al.Deep clustering for unsupervised learning of visual features[C]//Proceedings of the European Conference on Computer Vision.Munich:ECCV,2018:132-149.
[7] COATES A,NG A Y.Learning feature representations with K-means[J].Neural Networks:Tricks of the Trade,2012:561-580.
[8] SHARIF RAZAVIAN A,AZIZPOUR H,SULLIVAN J,et al.CNN features off-the-shelf:An astounding baseline for recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops.Columbus:ICCV,2014:806-813.
[9] BOJANOWSKI P, JOULIN A.Unsupervised learning by predicting noise[C]//International Conference on Machine Learning.Sydney:PMLR,2017:517-526.
[10] WENG Y,ZHANG N,YANG X.Improved density peak clustering based on information entropy for ancient character images[J].IEEE Access,2019,7:81691-81700.
[11] 陈扬,王金亮,夏炜,等.基于特征自动提取的足迹图像聚类方法[J].计算机科学,2021,48(S1):255-259.
[12] ZHAO H,CHU H,ZHANG Y,et al.Improvement of ancient shui character recognition model based on convolutional neural network[J].IEEE Access,2020,8:33080-33087.
[13] WANG X,GUPTA A.Unsupervised learning of visual representations using videos[C]//Proceedings of the IEEE International Conference on Computer Vision.Santiago:ICCV,2015:2794-2802.
[14] 魏银华.基于Python的古汉语文本聚类应用研究[D].大连:大连理工大学,2018.
[15] 李丁园,李晓杰.基于多尺度残差卷积自编码器的图像聚类方法[J].吉林大学学报(信息科学版),2022,40(4):684-687.
相似文献/References:
[1]郭洪壮,金小峰.基于HRCenterNet模型改进的朝鲜语古籍文字检测方法[J].延边大学学报(自然科学版),2022,(03):235.
GUO Hongzhuang,JIN Xiaofeng.Korean ancient books character detection method based on improved HRCenterNet model[J].Journal of Yanbian University,2022,(02):235.
备注/Memo
收稿日期: 2023 03 20
基金项目: 延边大学外国语言文学世界一流学科建设项目(18YLPY14);国家社会科学基金重大项目(18ZDA306);延边大学应用基础研究项目(延大科合字(2021)第2号)
第一作者: 刘晓童(1998—),女,硕士研究生,研究方向为计算机视觉.
通信作者: 金小峰(1970—),男(朝鲜族),硕士,教授,研究方向为语音信息处理、计算机视觉.