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

Optics and Precision Engineering, Volume. 33, Issue 12, 1955(2025)

MDAT：Multi-dimensional aggregation transformer for image super-resolution reconstruction

Qingjiang CHEN and Pengmin CHEN*
Author Affiliations
  • School of Science， Xi'an University of Architecture and Technology， Xi'an710055， China
    • show less
    AbstractGet PDF(in Chinese)
    Figures&Tables (18)
    Equations (29)
    References (16)
    Tools
    Paper Information
    Topics
    Figures & Tables(18)
    Network structure of MDAT

    Fig. 1. Network structure of MDAT

    Download full size
    View in Article
    Structure of the spatial transformer module and channel transformer module

    Fig. 2. Structure of the spatial transformer module and channel transformer module

    Download full size
    View in Article
    Structure of the spatial-channel interaction module

    Fig. 3. Structure of the spatial-channel interaction module

    Download full size
    View in Article
    Structure of the feature reuse transformer module

    Fig. 4. Structure of the feature reuse transformer module

    Download full size
    View in Article
    Performance comparison of different algorithms

    Fig. 5. Performance comparison of different algorithms

    Download full size
    View in Article
    Visualization of ×4SR task for different algorithms

    Fig. 6. Visualization of ×4SR task for different algorithms

    Download full size
    View in Article
    Visualization of ablation study on MSIMM

    Fig. 7. Visualization of ablation study on MSIMM

    Download full size
    View in Article
    Visualization of ablation study on SCIM

    Fig. 8. Visualization of ablation study on SCIM

    Download full size
    View in Article
    Visualization of ablation study on FRTM

    Fig. 9. Visualization of ablation study on FRTM

    Download full size
    View in Article
    Visualization of ablation study on core modules

    Fig. 10. Visualization of ablation study on core modules

    Download full size
    View in Article

    • Table 1. Quantitative evaluation of ×2 SR task on benchmark datasets

      View table
      View in Article

      Table 1. Quantitative evaluation of ×2 SR task on benchmark datasets

      方法模型复杂度Set5Set14BSD100Urban100Manga109
      Params/MACsPSNR/SSIMPSNR/SSIMPSNR/SSIMPSNR/SSIMPSNR/SSIM
      Bicubic-33.66/0.929 930.24/0.868 829.56/0.843 126.88/0.840 330.80/0.933 9
      SRCNN（2016）8/5336.66/0.954 232.45/0.906 731.36/0.887 929.50/0.894 635.60/0.966 3
      FSRCNN（2016）13/637.00/0.955 832.63/0.908 831.53/0.892 029.88/0.902 036.67/0.971 0
      CARN（2018）1 592/222.837.76/0.959 033.52/0.916 632.09/0.897 831.92/0.925 638.36/0.976 5
      SwinIR-L（2021）910/244.438.14/0.961 133.86/0.920 632.31/0.901 232.76/0.934 039.12/0.978 3
      DAT-L（2023）763/173.038.25/0.961 534.01/0.921 532.35/0.902 132.91/0.934 839.51/0.979 0
      Omni-SR（2023）772/194.538.22/0.961 333.98/0.921 032.36/0.902 033.05/0.936 339.28/0.978 4
      SwinIR-NG（2023）1 181/274.138.17/0.961 233.94/0.920 532.31/0.901 332.78/0.934 039.20/0.978 1
      CRAFT（2024）737/176.638.23/0.961 533.92/0.921 132.33/0.901 632.86/0.934 339.39/0.978 6
      MambaIR-light（2024）1 363/278.938.16/0.961 034.00/0.921 232.34/0.901 732.92/0.935 639.31/0.977 9
      SRConvNet-L（2025）885/16038.14/0.961 033.81/0.919 932.28/0.901 032.59/0.932 139.22/0.977 9
      HiT-SNG（2025）1 013/213.938.21/0.961 234.00/0.921 732.35/0.902 033.01/0.936 039.32/0.978 2
      Ours997/289.438.29/0.961 634.11/0.921 932.39/0.902 433.16/0.936 939.60/0.979 1

    • Table 2. Quantitative evaluation of ×3 SR task on benchmark datasets

      View table
      View in Article

      Table 2. Quantitative evaluation of ×3 SR task on benchmark datasets

      方法模型复杂度Set5Set14BSD100Urban100Manga109
      Params/MACsPSNR/SSIMPSNR/SSIMPSNR/SSIMPSNR/SSIMPSNR/SSIM
      Bicubic-30.39/0.868 227.55/0.774 227.21/0.738 524.46/0.734 926.95/0.855 6
      SRCNN（2016）8/5332.75/0.909 029.30/0.821 528.41/0.786 326.24/0.798 930.48/0.911 7
      FSRCNN（2016）13/533.18/0.914 029.37/0.824 028.53/0.791 026.43/0.808 031.10/0.921 0
      CARN（2018）1 592/119.034.29/0.925 530.29/0.840 729.06/0.803 428.06/0.849 333.50/0.944 0
      SwinIR-L（2021）918/110.834.62/0.928 930.54/0.846 329.20/0.808 228.66/0.862 433.98/0.947 8
      DAT-L（2023）947/96.734.75/0.929 930.63/0.847 529.30/0.810 628.90/0.866 834.55/0.950 2
      Omni-SR（2023）780/88.434.70/0.929 430.57/0.846 929.28/0.809 428.84/0.865 634.22/0.948 7
      SwinIR-NG（2023）1 190/114.134.64/0.929 330.58/0.847 129.24/0.809 028.75/0.863 934.22/0.948 8
      CRAFT（2024）744/78.434.71/0.929 530.61/0.846 929.24/0.809 328.77/0.863 534.29/0.949 1
      MambaIR-light（2024）1 371/124.634.72/0.929 630.63/0.847 529.29/0.809 929.00/0.868 934.39/0.949 0
      SRConvNet-L（2025）906/7434.59/0.928 830.50/0.845 529.22/0.808 128.56/0.860 034.17/0.947 9
      HiT-SNG（2025）1 021/99.534.74/0.929 730.62/0.847 429.26/0.810 028.91/0.867 134.38/0.949 5
      Ours1 007/129.834.76/0.930 030.65/0.847 629.32/0.810 729.00/0.868 334.64/0.950 5

    • Table 3. Quantitative evaluation of ×4 SR task on benchmark datasets

      View table
      View in Article

      Table 3. Quantitative evaluation of ×4 SR task on benchmark datasets

      方法模型复杂度Set5Set14BSD100Urban100Manga109
      Params/MACsPSNR/SSIMPSNR/SSIMPSNR/SSIMPSNR/SSIMPSNR/SSIM
      Bicubic-28.42/0.810 426.00/0.702 725.96/0.667 523.14/0.657 724.89/0.786 6
      SRCNN（2016）8/5330.48/0.862 627.50/0.751 326.90/0.710 124.52/0.722 127.58/0.855 5
      FSRCNN（2016）13/4.630.72/0.866 027.61/0.755 026.98/0.715 024.62/0.728 027.90/0.861 0
      CARN（2018）1 592/91.032.13/0.893 728.60/0.780 627.58/0.734 926.07/0.783 730.47/0.908 4
      SwinIR-L（2021）930/63.632.44/0.897 628.77/0.785 827.69/0.740 626.47/0.798 030.92/0.915 1
      DAT-L（2023）910/78.432.57/0.899 228.87/0.788 027.75/0.743 126.65/0.803 531.37/0.917 9
      Omni-SR（2023）792/50.932.49/0.898 828.78/0.785 927.71/0.741 526.64/0.801 831.02/0.915 1
      SwinIR-NG（2023）1 201/64.432.44/0.898 028.83/0.787 027.73/0.741 826.61/0.801 031.09/0.916 1
      CRAFT（2024）753/46.532.52/0.898 928.85/0.787 227.72/0.741 826.56/0.799 531.18/0.916 8
      MambaIR-light（2024）1 280/42.932.36/0.898 428.53/0.789 527.78/0.744 626.68/0.805 731.17/0.917 6
      SRConvNet-L（2025）902/4532.44/0.897 628.77/0.785 727.69/0.740 226.47/0.797 030.96/0.913 9
      HiT-SNG（2025）1 032/57.732.55/0.899 128.83/0.787 327.74/0.742 626.75/0.805 331.24/0.917 6
      Ours1 023/73.832.60/0.899 728.91/0.788 527.77/0.743 426.77/0.806 331.38/0.917 2

    • Table 4. Per-frame average inference time

      View table
      View in Article

      Table 4. Per-frame average inference time

      放大倍数输出分辨率单帧图像平均推理时间/s
      SwinIR-LMDAT（Ours）
      ×21 280×7201.164 71.190 0
      ×30.510 90.501 8
      ×40.282 30.264 2

    • Table 5. Ablation study of MSIMM

      View table
      View in Article

      Table 5. Ablation study of MSIMM

      实验对象模型复杂度Urban100Manga109
      Params/MACsPSNR/SSIMPSNR/SSIM
      Re-Conv3×3871/197.433.00/0.935 739.50/0.979 0
      Re-3*Conv3×3871/200.432.00/0.935 639.50/0.979 0
      W-MSIMM885/203.533.03/0.935 939.53/0.979 1

    • Table 6. Ablation study of SCIM

      View table
      View in Article

      Table 6. Ablation study of SCIM

      实验对象模型复杂度Urban100Manga109
      Params/MACsPSNR/SSIMPSNR/SSIM
      WO-SCIM885/203.533.03/0.935 939.53/0.979 1
      W-SI891/205.032.95/0.935 339.51/0.979 0
      W-CI912/203.832.91/0.934 839.49/0.979 0
      W-SCIM919/205.333.11/0.936 739.58/0.979 1

    • Table 7. Ablation study of FRTM

      View table
      View in Article

      Table 7. Ablation study of FRTM

      实验对象模型复杂度Urban100Manga109
      Params/MACsPSNR/SSIMPSNR/SSIM
      WO-FRTM919/205.333.11/0.936 739.58/0.979 1
      W-Res919/205.332.94/0.935 039.50/0.979 0
      W-FRTM997/289.433.16/0.936 939.60/0.979 1

    • Table 8. Ablation study of Core Modules

      View table
      View in Article

      Table 8. Ablation study of Core Modules

      MSIMMSCIMFRTM模型复杂度Urban100Manga109
      Params/MACsPSNR/SSIMPSNR/SSIM
      ×××871/197.433.00/0.935 739.50/0.979 0
      √××885/203.533.03/0.935 939.53/0.979 1
      √√×919/205.333.11/0.936 739.58/0.979 1
      √√√997/289.433.16/0.936 939.60/0.979 1
    Tools

    Get Citation

    Copy Citation Text

    Qingjiang CHEN, Pengmin CHEN. MDAT：Multi-dimensional aggregation transformer for image super-resolution reconstruction[J]. Optics and Precision Engineering, 2025, 33(12): 1955

    Download Citation

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

    Category:

    Received: Feb. 13, 2025

    Accepted: --

    Published Online: Aug. 15, 2025

    The Author Email:

    DOI:10.37188/OPE.20253312.1955

    Topics

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