Laser & Optoelectronics Progress, Volume. 61, Issue 4, 0411003(2024)

Lightweight Image Super-Resolution Based on Shuffle Group Convolution and Sparse Global Attention

Xiang Li and Juan Zhang*
Author Affiliations
  • School of Electrical and Electronic Engineering, Shanghai University of Technology, Shanghai 201620, China
  • show less
    Figures & Tables(13)
    Overall network structure
    Schematic diagram of the comparison between the operation process of shuffle group convolution and conventional convolution
    Structure diagram of SGA mechanism
    Example of reconstruction results to img-49
    Example of reconstruction results to img-78
    Example of reconstruction results to real dataset
    • Table 1. Comparative tests on different test sets when the scale factor is 2

      View table

      Table 1. Comparative tests on different test sets when the scale factor is 2

      MethodSet5Set14Urban100B100
      PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
      EDSR38.110.960233.920.919532.930.935132.320.9013
      RDN38.240.961434.010.921232.890.935332.340.9017
      RCAN38.270.961434.120.921633.340.938432.410.9027
      SAN38.310.962034.070.921333.100.937032.420.9028
      HAN38.270.961434.160.921733.350.938532.410.9027
      SwinIR38.350.962034.140.922733.400.939332.440.9030
      ENLCN38.370.961834.170.922933.560.939832.490.9032
      SSIR38.370.962034.160.922733.380.938732.580.9039
    • Table 2. Comparative tests on different test sets when the scale factor is 3

      View table

      Table 2. Comparative tests on different test sets when the scale factor is 3

      MethodSet5Set14Urban100B100
      PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
      EDSR34.650.928030.520.846228.800.865329.250.8093
      RDN34.710.929630.570.846828.800.865329.260.8093
      RCAN34.740.929930.650.848228.090.870229.320.8111
      SAN34.750.930030.590.847628.930.867129.330.8112
      HAN34.750.929930.670.848329.100.870529.320.8110
      SwinIR34.890.931230.770.850329.290.874429.370.8124
      ENLCN
      SSIR34.870.931230.830.850529.330.875729.370.8125
    • Table 3. Comparative tests on different test sets when the scale factor is 4

      View table

      Table 3. Comparative tests on different test sets when the scale factor is 4

      MethodSet5Set14Urban100B100
      PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
      EDSR32.460.896828.800.787626.640.803327.710.7420
      RDN32.470.899028.810.787126.610.802827.720.7419
      RCAN32.630.900228.870.788926.820.808727.770.7436
      SAN32.640.900328.920.788826.790.806827.780.7436
      HAN32.640.900228.900.789026.850.809427.800.7442
      SwinIR32.720.902128.940.791427.070.816427.830.7459
      ENLCN32.670.900428.940.789227.120.818427.820.7452
      SSIR32.740.902128.940.791127.120.816927.910.7465
    • Table 4. Comparison of reference quantity, calculation quantity and running time of each method

      View table

      Table 4. Comparison of reference quantity, calculation quantity and running time of each method

      MethodScale factor is 2Scale factor is 3Scale factor is 4
      Parameter quantity /×106Flops /×109Latency /msParameter quantity /×106Flops /×109Latency /msParameter quantity /×106Flops /×109Latency /ms
      EDSR40.739387114343.86447057343.092895360
      RDN22.12509884622.30228240622.271310243
      RCAN15.443530353015.62158636715.59918223
      SAN15.86305015.89162015.86937
      HAN63.6014551227864.346534101464.193776628
      SwinIR11.752301291311.931026123811.90584645
      SSIR9.24198712459.1810044279.18532332
    • Table 5. Ablation experiments for the shuffle group convolution

      View table

      Table 5. Ablation experiments for the shuffle group convolution

      MethodParameter quantity /×106Flops /×109PSNRSSIM
      Shuffle group convolution9.24189738.370.9620
      3×3 convolution14.86301538.410.9621
    • Table 6. Ablation experiments for the SGA mechanism

      View table

      Table 6. Ablation experiments for the SGA mechanism

      MethodPSNRSSIMParameter quantity /×106Flops /×109
      LFEB+SGA38.370.96209.241897
      LFEB+MSA38.290.961412.682258
      LEFB38.100.96016.721372
    • Table 7. Ablation experiment on the number of MFRM

      View table

      Table 7. Ablation experiment on the number of MFRM

      ModulePSNRSSIM
      Interpolation38.050.9560
      Dconv38.180.9613
      MFRM38.370.9620
    Tools

    Get Citation

    Copy Citation Text

    Xiang Li, Juan Zhang. Lightweight Image Super-Resolution Based on Shuffle Group Convolution and Sparse Global Attention[J]. Laser & Optoelectronics Progress, 2024, 61(4): 0411003

    Download Citation

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

    Category: Imaging Systems

    Received: Apr. 10, 2023

    Accepted: May. 29, 2023

    Published Online: Feb. 6, 2024

    The Author Email: Juan Zhang (zhang-j@foxmail.com)

    DOI:10.3788/LOP231061

    CSTR:32186.14.LOP231061

    Topics