Journals Articles Conferences News About CLP
Submit a paper Email Alert
Search
Search
Articles
Journals
News
Advanced Search
Top Searches
2D MaterialsTransformation opticsQuantum PhotonicsSmall lasersSemiconductor UV PhotonicsSupersymmetric microring laser arrays

Laser & Optoelectronics Progress, Volume. 61, Issue 10, 1037009(2024)

Global-Sampling Spatial-Attention Module and its Application in Image Classification and Small Object Detection and Recognition

Jingyu Lu1,2,3, Haiyang Zhang1,2,3、*, Wenxin Wang1,2,3, and Changming Zhao1,2,3
Author Affiliations
  • 1School of Optics and Photonics, Beijing Institute of Technology, Beijing 100081
  • 2Key Laboratory of Optoelectronic Imaging Technology and Systems, Ministry of Education, Beijing 100081
  • 3Key Laboratory of Information Photonics Technology, Ministry of Industry and Information Technology, Beijing 100081
    • show less
    AbstractGet PDF(in Chinese)
    Figures&Tables (21)
    Equations (3)
    References (22)
    Cited By (1)
    Tools
    Paper Information
    Topics
    Figures & Tables(21)
    BAM architecture[1] and CBAM architecture[11]

    Fig. 1. BAM architecture[1] and CBAM architecture[11]

    Download full size
    Architecture of non-local neural network attention module[13]

    Fig. 2. Architecture of non-local neural network attention module[13]

    Download full size
    Architecture of 3×3 deformable convolution module[12]

    Fig. 3. Architecture of 3×3 deformable convolution module[12]

    Download full size
    Detailed architecture of the GSSAM

    Fig. 4. Detailed architecture of the GSSAM

    Download full size
    Dimension reduction of input feature map

    Fig. 5. Dimension reduction of input feature map

    Download full size
    Schematic of benchmark selection process

    Fig. 6. Schematic of benchmark selection process

    Download full size
    Process of obtaining attention map by difference comparison

    Fig. 7. Process of obtaining attention map by difference comparison

    Download full size
    Loss and accuracy changes on training set. (a) Loss change; (b) accuracy change

    Fig. 8. Loss and accuracy changes on training set. (a) Loss change; (b) accuracy change

    Download full size
    Loss change on validation set. (a) Original loss change; (b) loss fitting; (c) local comparison

    Fig. 9. Loss change on validation set. (a) Original loss change; (b) loss fitting; (c) local comparison

    Download full size
    Accuracy change on validation set. (a) Original accuracy change; (b) accuracy fitting; (c) local comparison

    Fig. 10. Accuracy change on validation set. (a) Original accuracy change; (b) accuracy fitting; (c) local comparison

    Download full size
    Comparison of accuracy fitting curves of different networks. (a) VGG19 fitting curves; (b) VGG19 local fitting curves; (c) ResNet50 fitting curves; (d) ResNet50 local fitting curves

    Fig. 11. Comparison of accuracy fitting curves of different networks. (a) VGG19 fitting curves; (b) VGG19 local fitting curves; (c) ResNet50 fitting curves; (d) ResNet50 local fitting curves

    Download full size
    Comparison of loss fitting curves of different networks. (a) VGG19 fitting curves; (b) VGG19 local fitting curves; (c) ResNet50 fitting curves; (d) ResNet50 local fitting curves

    Fig. 12. Comparison of loss fitting curves of different networks. (a) VGG19 fitting curves; (b) VGG19 local fitting curves; (c) ResNet50 fitting curves; (d) ResNet50 local fitting curves

    Download full size
    Comparison of Grad-CAM's visualization results

    Fig. 13. Comparison of Grad-CAM's visualization results

    Download full size
    Schematic of the role of GSSAM in small target images

    Fig. 14. Schematic of the role of GSSAM in small target images

    Download full size
    Schematic of GSSAM embedding YOLOx

    Fig. 15. Schematic of GSSAM embedding YOLOx

    Download full size
    Partial “low slow small” UAV images

    Fig. 16. Partial “low slow small” UAV images

    Download full size
    Schematic of residual GSSAM structure

    Fig. 17. Schematic of residual GSSAM structure

    Download full size
    Schematic of module structure without sampling and comparison

    Fig. 18. Schematic of module structure without sampling and comparison

    Download full size

    • Table 1. Classification results of different networks for images on some ImageNet-1K data sets

      View table

      Table 1. Classification results of different networks for images on some ImageNet-1K data sets

      ModelParamsFLOPs /109Accuracy /%
      ResNet3425，856，58611.2680.00
      ResNet34+CBAM25，908，78811.2676.80
      ResNet34+GSSAM25，864，72811.2781.00
      ResNet5023，528，5225.1978.20
      ResNet50+CBAM24，225，8445.1980.80
      ResNet50+GSSAM23，551，2925.2179.80
      ResNet10142，520，65010.0579.60
      ResNet101+CBAM43，217，97210.0580.00
      ResNet101+GSSAM42，543，45610.0779.40
      ResNet15258，164，29814.8577.80
      ResNet152+CBAM58，861，62014.8577.20
      ResNet152+GSSAM58，187，06814.8779.40
      DarkNet5340，613，0348.9980.00
      DarkNet53+CBAM40，788，1169.0081.20
      DarkNet53+GSSAM40，624，7429.0181.60
      VGG19139，622，21823.0282.60
      VGG19+CBAM139，698，99623.0381.80
      VGG19+GSSAM139，632，92023.0484.00

    • Table 2. Statistical of experimental results

      View table

      Table 2. Statistical of experimental results

      ModelTime /msmAP50∶95 /%
      Origin YOLOx20.9658.67
      YOLOx-GSSAM Sigmoidposition 122.4660.64
      Position 222.4760.50
      YOLOx-GSSAM standardizationposition 124.9659.50
      Position 224.4358.96
      YOLOx-GSSAM Sigmoid Resposition 122.3161.44
      position 221.9961.01

    • Table 3. Verification for sampling and comparison effects

      View table

      Table 3. Verification for sampling and comparison effects

      ParameterGSSAMGSSAM ResConvConv Res
      position 1position 2position 1position 2position 1position 2position 1position 2
      mAP50∶95 /%60.6460.5061.4461.0159.5659.4960.0160.06
    Tools

    Get Citation

    Copy Citation Text

    Jingyu Lu, Haiyang Zhang, Wenxin Wang, Changming Zhao. Global-Sampling Spatial-Attention Module and its Application in Image Classification and Small Object Detection and Recognition[J]. Laser & Optoelectronics Progress, 2024, 61(10): 1037009

    Download Citation

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

    Category: Digital Image Processing

    Received: Aug. 18, 2023

    Accepted: Oct. 9, 2023

    Published Online: Apr. 29, 2024

    The Author Email: Haiyang Zhang (ocean@bit.edu.cn)

    DOI:10.3788/LOP231933

    CSTR:32186.14.LOP231933

    Topics

    laser devices and laser physics
    Lasers and Laser Optics
    Laser physics
    laser manufacturing
    Instrumentation, Measurement and Metrology