• Laser & Optoelectronics Progress
  • Vol. 58, Issue 6, 610016 (2021)
Fu Xingwu1, Lü Mingming1、2、*, Liu Wanjun1, and Wei Xian2
Author Affiliations
  • 1College of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
  • 2Quanzhou Institute of Equipment Manufacturing Haixi Institutes, Chinese Academy of Sciences, Quanzhou, Fujian 362200, China
  • show less
    DOI: 10.3788/LOP202158.0610016 Cite this Article Set citation alerts
    Fu Xingwu, Lü Mingming, Liu Wanjun, Wei Xian. Structured Deep Discriminant Embedded Coding Network for Image Clustering[J]. Laser & Optoelectronics Progress, 2021, 58(6): 610016 Copy Citation Text show less
    General network architecture of deep clustering algorithm based on AE
    Fig. 1. General network architecture of deep clustering algorithm based on AE
    General network architecture of deep clustering algorithm based on VAE
    Fig. 2. General network architecture of deep clustering algorithm based on VAE
    Overall network framework of the SDDECC algorithm
    Fig. 3. Overall network framework of the SDDECC algorithm
    Local mutual information estimation network
    Fig. 4. Local mutual information estimation network
    Some samples in MNIST and Fashion-MNIST datasets. (a) MNIST dataset; (b) Fashion-MNIST dataset
    Fig. 5. Some samples in MNIST and Fashion-MNIST datasets. (a) MNIST dataset; (b) Fashion-MNIST dataset
    Distribution visualization of embedded subspaces of different strategies on the MNIST dataset. (a) ConvAE; (b) ConvAE+MI; (c) ConvAE+GCN; (d) ConvAE+MI+GCN
    Fig. 6. Distribution visualization of embedded subspaces of different strategies on the MNIST dataset. (a) ConvAE; (b) ConvAE+MI; (c) ConvAE+GCN; (d) ConvAE+MI+GCN
    Effect of different combinations of parameter λ1 and λ2 on the ACC and NMI on the Fashion-MNIST dataset. (a) Effect on ACC; (b) effect on NMI
    Fig. 7. Effect of different combinations of parameter λ1 and λ2 on the ACC and NMI on the Fashion-MNIST dataset. (a) Effect on ACC; (b) effect on NMI
    DatasetNumber of samplesNumber of classesDimension
    USPS9298101×16×16
    MNIST70000101×28×28
    Fashion-MNIST70000101×28×28
    Table 1. Dataset introduction
    DatasetEncoder-1/Decoder-4Encoder-2/Decoder-3Encoder-3/Decoder-2Encoder-4/Decoder-1
    USPS3×3×163×3×32
    MNIST3×3×163×3×163×3×323×3×32
    Fashion-MNIST3×3×163×3×163×3×323×3×32
    Table 2. Number of channels and core size of autoencoder network
    AlgorithmUSPSMNISTFashion-MNIST
    ACCNMIACCNMIACCNMI
    K-means0.66820.62700.53220.50040.47420.5120
    AE+K-means0.69310.66200.80760.73030.58530.6142
    DEC0.74080.75290.86550.83720.51800.5462
    IDEC0.76050.78460.88060.86720.52910.5570
    Deepcluster0.56230.54030.79710.66150.54220.5100
    SDCN0.77890.79260.85300.84270.57800.6047
    SDDECC0.79860.81420.90220.89580.61710.6306
    Table 3. Clustering results of different clustering algorithms on three datasets
    MethodUSPSMNISTFashion-MNIST
    ACCNMIACCNMIACCNMI
    ConvAE0.69810.65190.77620.74500.54620.5563
    ConvAE+MI0.78530.74420.83500.80230.59220.6091
    ConvAE+GCN0.78220.78750.85740.84490.58430.6167
    ConvAE+MI+GCN0.79860.81420.90220.89580.61710.6306
    Table 4. Impact of different strategies on clustering performance
    Fu Xingwu, Lü Mingming, Liu Wanjun, Wei Xian. Structured Deep Discriminant Embedded Coding Network for Image Clustering[J]. Laser & Optoelectronics Progress, 2021, 58(6): 610016
    Download Citation