Optics and Precision Engineering, Volume. 30, Issue 18, 2241(2022)

Unsupervised representation learning for cultural relics based on local-global bidirectional reasoning

Jie LIU1...2, Guohua GENG1,2,*, Yu TIAN1,2, Yi WANG1,2, Yangyang LIU1,2, and Mingquan ZHOU12 |Show fewer author(s)
Author Affiliations
  • 1National and Local Joint Engineering Research Center for Cultural Heritage Digitization, Northwest University, Xi'an7027, China
  • 2College of Information Science and Technology, Northwest University, Xi'an71017, China
  • show less
    Figures & Tables(17)
    Examples of the Terracotta Warriors fragments datasets
    Approach of the extend dataset
    Local-global bidirectional reasoning
    Overall network framework
    Multi-scale shell convolution block
    Illustration of Triplet loss and N-pair loss
    Training loss curves for different decoders
    • Table 1. Encoder network parameters

      View table
      View in Article

      Table 1. Encoder network parameters

      LayerMjsjXikiMLP1FprevMLP2 out sizeShellconv out sizeMLP3 Outsize
      1st51232132[32, 64]-(512,32,64)(512,1,64)(512,1,256)
      264(512,64,64)(512,1,64)
      4128(512,128,64)(512,1,128)
      2nd12816116[32, 64]256(128,16,320)(128,1,128)(128,1,896)
      232(128,32,320)(128,1,256)
      464(128,64,320)(128,1,512)
    • Table 2. Classification accuracies of different methods on Terracotta Warrior fragments dataset.

      View table
      View in Article

      Table 2. Classification accuracies of different methods on Terracotta Warrior fragments dataset.

      MethodInput data typeDeep modelSupervisedOverall accuracy
      Template-based method26PFT87.64%
      PointNet11PTT88.93%
      Dual-modal7P, GTT91.41%
      ShellNet13PTT92.44%
      AMS-Net17PTT95.68%
      Foldingnet20PTF81.91%
      Proposed methodPTF93.33%
    • Table 3. Classification performance of different models

      View table
      View in Article

      Table 3. Classification performance of different models

      MethodPRF
      ShellNet1392.55%92.61%92.49%
      AMS-Net1795.68%95.71%95.65%
      Foldingnet2082.66%81.84%82.29%
      Proposed method93.33%93.34%93.25%
    • Table 4. Classification accuracies of the four classes

      View table
      View in Article

      Table 4. Classification accuracies of the four classes

      MethodArmBodyHeadLeg
      Dual-modal7(G)77.75%92.75%91.50%76.25%
      Dual-modal7(P)82.51%96.45%92.36%84.41%
      Dual-modal7(G+P)87.55%87.55%94.37%88.41%
      AMS-Net1792.40%98.10%98.00%94.20%
      Proposed method84.88%97.08%94.97%91.76%
    • Table 5. Effect of the encoder on classification accuracies

      View table
      View in Article

      Table 5. Effect of the encoder on classification accuracies

      EncoderAccuracy
      Single-scale90.93%
      Multi-scale93.33%
    • Table 6. Effect of the decoder on classification accuracies

      View table
      View in Article

      Table 6. Effect of the decoder on classification accuracies

      DecoderAccuracy
      MLPs91.45%
      Folding-based93.33%
    • Table 7. Effect of the loss function on classification accuracies

      View table
      View in Article

      Table 7. Effect of the loss function on classification accuracies

      lossAccuracy
      Lrec89.47%
      Lim-N+Lrec93.33%
    • Table 8. Results for different noise points

      View table
      View in Article

      Table 8. Results for different noise points

      Noise NumberAccuracy
      192.13%
      1088.93%
      5075.62%
      10072.77%
    • Table 9. Comparisons of the classification accuracy of our method against the unsupervised learning methods on ModelNet40

      View table
      View in Article

      Table 9. Comparisons of the classification accuracy of our method against the unsupervised learning methods on ModelNet40

      Unsupervised MethodAccuracy
      FoldingNet2088.40%
      l-GAN (M40)1987.27%
      l-GAN1985.70%
      Multi-Task2389.10%
      L2G Auto-encoder2290.64%
      Proposed method92.02%
    • Table 10. Comparisons of the classification accuracy of our method against the supervised learning methods on ModelNet40

      View table
      View in Article

      Table 10. Comparisons of the classification accuracy of our method against the supervised learning methods on ModelNet40

      Supervised methodAccuracy
      PointNet1189.20%
      PointNet++1290.70%
      ShellNet1393.10%
      SO-Net1490.90%
      Proposed method92.02%
    Tools

    Get Citation

    Copy Citation Text

    Jie LIU, Guohua GENG, Yu TIAN, Yi WANG, Yangyang LIU, Mingquan ZHOU. Unsupervised representation learning for cultural relics based on local-global bidirectional reasoning[J]. Optics and Precision Engineering, 2022, 30(18): 2241

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Information Sciences

    Received: Mar. 27, 2022

    Accepted: --

    Published Online: Oct. 27, 2022

    The Author Email: GENG Guohua (ghgeng@nwu.edu.cn)

    DOI:10.37188/OPE.20223018.2241

    Topics