基于无监督学习视觉特征的深度聚类方法
作者:
作者单位:

1.河北大学数学与信息科学学院河北省机器学习与计算智能重点实验室,保定 071002;2.北京师范大学珠海分校应用数学学院,珠海 519087

作者简介:

通讯作者:

陈俊芬,女,副教授,E-mail:chenjunfen2010@126.com。

中图分类号:

TP391

基金项目:

河北省引进留学人员基金(C20200302)资助项目;河北省科技重点研发计划(19210310D)资助项目;广东省自然科学基金(2018A0303130026)资助项目。


Deep Clustering Method Based on Unsupervised Visual Features Learning
Author:
Affiliation:

1.Hebei Key Laboratory of Machine Learning and Computational Intelligence, College of Mathematics and Information Sciences, Hebei University, Baoding 071002, China;2.School of Applied Mathematics, Beijing Normal University (Zhuhai), Zhuhai 519087, China

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    基于自编码器的特征提取技术广泛应用于图像聚类分析,在较简单的图像集上取得了令人满意的聚类结果,但自编码器的特征表示能力有限,很难捕捉到复杂低质图像的局部特征。本文提出一种基于非对称结构卷积自编码器(Convolutional auto-encoder with an asymmetric structure, ASCAE)的学习视觉特征的深度聚类方法,其中非对称结构的卷积自编码器用于学习特征表示,然后使用K-means算法对特征数据进行聚类分析。为进一步提高特征表示能力,ASCAE方法的网络采用变步长的卷积层和全连接的重构误差正则约束网络的重构误差。在7个公开图像集上的实验结果表明该网络有很好的特征表示能力,并且使得K-means算法能提供很好的聚类结果。在COIL-20和MNIST图像集上,聚类方法ASCAE的聚类精度分别为0.754和0.918,优于同类型的4种深度聚类方法(AEC、IEC、DEC和DEN)。

    Abstract:

    Despite recent progress in features extraction using deep conventional auto-encoders, which have greatly benefited image clustering analysis with the satisfying clustering results on several simple image-datasets. However, the representation ability of the conventional auto-encoders is quite limit when they capture local features of complex and low-quality images. Therefore, a novel deep clustering method combining with visual features learning is proposed,i.e., the convolutional auto-encoder with an asymmetric structure(ASCAE), in which an asymmetric convolutional auto-encoder is used to learn feature representation, and then K-means algorithm performs clustering analysis for these learned features. To further improve the representation suitability for downstream clustering tasks, ASCAE method adopts strided convolution layers, and minimizes the reconstruction error of whole network regularized by L2 error between the left and right fully connected layers. Experimental results on seven public image-datasets illustrate that the network of ASCAE method usually offers better feature representation and brings promising clustering performance presented by K-means algorithm. Clustering accuracy of ASCAE method is 0.754 and 0.918 on databases COIL-20 and MNIST, respectively, which is better than four depth clustering methods of the same type (AEC, IEC, DEC and DEN).

    参考文献
    相似文献
    引证文献
引用本文

陈俊芬,赵佳成,翟俊海,李艳.基于无监督学习视觉特征的深度聚类方法[J].南京航空航天大学学报,2021,53(5):718-725

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2020-02-15
  • 最后修改日期:2021-01-17
  • 录用日期:
  • 在线发布日期: 2021-11-02
  • 出版日期:
您是第位访问者
南京航空航天大学学报 ® 2024 版权所有
技术支持:北京勤云科技发展有限公司