吉林大学学报(信息科学版) ›› 2022, Vol. 40 ›› Issue (4): 684-687.

• • 上一篇    下一篇

基于多尺度残差卷积自编码器的图像聚类方法

李丁园1 , 李晓杰2   

  1. 1. 中国电子科技集团公司电子科学研究院, 北京 100041; 2. 内蒙古机电职业技术学院 电气工程系, 呼和浩特 010070
  • 收稿日期:2022-04-29 出版日期:2022-08-16 发布日期:2022-08-17
  • 作者简介:李丁园(1987— ), 女, 山东蓬莱人, 中国电子科技集团公司电子科学研究院工程师, 博士, 主要从事人工智能与大数据 应用研究, (Tel)86-15801087029(E-mail)lidingyuan@cetc.com.cn。
  • 基金资助:
    国家自然科学基金资助项目(U19B2036; U20B2062)

Image Clustering Method Based on Multi-Scale Residual Convolutional Autoencoder

LI Dingyuan 1 , LI Xiaojie 2   

  1. 1. China Academic of Electronics and Information Technology, Beijing 100041, China; 2. Department of Electrical Engineering, Inner Mongolia Technical College of Mechanics and Electrics, Hohhot 010070, China
  • Received:2022-04-29 Online:2022-08-16 Published:2022-08-17

摘要: 对于图像的聚类, 现有方法在特征提取方面或难以选择合适的维度转换方法, 或提取的特征对图像特征的表达较弱且不够丰富, 对图像的聚类效果产生了较大影响, 导致了聚类精度较低。 为此, 提出一种基于多尺度残差卷积自编码器的图像聚类方法, 通过构建具有若干个含有残差连接的多尺度卷积模块, 获得中间层的高维特征表达, 并以此进行图像聚类。 实验结果表明, 在 MNIST 数据集上的聚类准确率为 82. 2% , ARI(Adjusted Rand Index) 值为 0. 781 0, NMI (Normalized Mutual Information) 值为 0. 853 2, 模型达到了较好的聚类效果。

关键词: 卷积自编码器; , 图像聚类; , 多尺度残差连接; , 深度学习

Abstract: For image clustering, the existing methods are either difficult to choose the appropriate dimension transformation method in feature extraction, or the extracted features are weak and not rich enough for the expression of image features, which have a great impact on the clustering effect of images and lead to low clustering accuracy. Therefore, an image clustering method based on multi-scale residual convolutional autoencoder is proposed. By constructing several multi-scale convolutional modules with residual connections, the high-dimensional feature expression of the middle layer is obtained, and the image is clustered based on these features. The clustering accuracy on MNIST data set is 82. 2% , ARI (Adjusted Rand Index) value is 0. 781 0 and NMI ( Normalized Mutual Information) value is 0. 853 2, indicating that the model has achieved good clustering effect.

Key words: convolutional autoencoder; , image clustering; , multi-scale residual connection; , deep learning


中图分类号: 

  • TP3. 05