• Optics and Precision Engineering
  • Vol. 32, Issue 6, 843 (2024)
Ying ZHOU1,2,*, Shenghu PEI1, Haiyong CHEN1,2, and Shibo XU1
Author Affiliations
  • 1School of Artificial Intelligence, Hebei University of Technology, Tianjin30030, China
  • 2China Hebei Control Engineering Research Center, Tianjin300130, China
  • show less
    DOI: 10.37188/OPE.20243206.0843 Cite this Article
    Ying ZHOU, Shenghu PEI, Haiyong CHEN, Shibo XU. Image super-resolution network based on multi-scale adaptive attention[J]. Optics and Precision Engineering, 2024, 32(6): 843 Copy Citation Text show less
    Network architecture of MAAN
    Fig. 1. Network architecture of MAAN
    Architecture of BU
    Fig. 2. Architecture of BU
    Architecture of ADB
    Fig. 3. Architecture of ADB
    Architecture of AWU
    Fig. 4. Architecture of AWU
    Architecture of MPIB
    Fig. 5. Architecture of MPIB
    Architecture of ADA
    Fig. 6. Architecture of ADA
    Comparison of the visual effect of different attention
    Fig. 7. Comparison of the visual effect of different attention
    Comparison of visual effect of each method with a scale factor of ×4
    Fig. 8. Comparison of visual effect of each method with a scale factor of ×4
    数量6789
    参数量/M7.368.529.6110.83
    PSNR38.0138.2038.3738.40
    Table 1. Ablation experiments of the numbers of MFFB
    结构Set5Set14BSD100Urban100
    PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
    结构132.350.896 328.660.782 127.450.727 426.580.799 8
    结构232.390.895 928.780.787 627.420.729 426.640.803 3
    结构332.430.899 128.850.798 227.540.731 926.590.802 8
    结构432.690.900 229.030.810 527.740.745 126.920.816 7
    结构532.760.901 829.110.811 827.830.746 327.030.819 5
    结构632.770.902 029.120.812 027.850.746 927.080.819 4
    完整的BU32.790.902 329.150.812 827.870.747 327.120.820 9
    Table 2. Results of ablation experiments of the structure of BU
    分支3×35×57×73×3+5×53×3+5×5+7×73×3+5×5+AWU
    参数量/M3.215.848.539.5019.999.69
    PSNR34.2634.3734.4334.8134.9534.91
    Table 3. Results of ablation experiments of the structure of ADB
    组数248
    参数量/M7.799.6114.55
    PSNR37.8238.3738.53
    Table 4. Ablation experiments of the groups of MPIB
    方法参数量/MSet5Set14BSD100Urban100
    PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
    基础网络9.6534.660.928 830.520.847 429.120.809 728.910.872 3
    +ECA9.6734.720.929 830.670.848 329.210.811 029.010.872 4
    +CBAM9.7334.780.930 130.700.848 729.260.811 429.090.873 1
    +BAM9.7834.760.929 930.720.849 329.220.811 529.180.873 6
    +ADA9.6934.930.931 730.800.850 929.400.813 629.370.875 3
    Table 5. Comparison of the evaluation metrics of different attention
    比例方法参数量/MSet5Set14BSD100Urban100
    PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
    ×3EDSR43.734.650.928 030.520.846 229.250.809 328.800.865 3
    RDN22.334.710.929 630.570.846 829.260.809 328.800.865 3
    E-GSCN15.934.690.929 430.590.847 129.280.809 928.870.866 9
    DRLN33.734.780.930 330.730.848 829.360.811 729.210.872 2
    SwinIR11.934.890.931 230.770.850 329.370.812 429.290.874 4
    MAAN9.6934.930.931 730.800.850 929.400.813 629.370.875 3
    Bicubic-28.430.802 226.100.693 625.970.651 723.140.659 9
    SRCNN0.0230.480.862 827.490.750 326.900.710 124.520.722 1
    VDSR0.6731.350.883 828.010.767 427.290.725 125.180.752 4
    DRRN0.3031.680.888 828.210.772 027.380.728 425.440.763 8
    A2F-L1.3732.320.896 428.670.783 927.620.737 936.320.793 1
    ×4JTF-SISR18.532.530.898 928.800.786 727.700.741 026.620.802 2
    EDSR43.132.460.896 828.800.787 627.710.742 026.640.803 3
    RDN22.632.470.899 028.810.787 127.720.741 926.610.802 8
    E-GSCN15.932.530.899 228.840.787 927.740.742 426.710.805 0
    DRLN34.032.630.900 228.940.790 027.830.744 426.980.811 9
    SwinIR11.932.750.902 128.940.791 427.830.745 927.070.816 4
    MAAN9.7232.790.902 329.150.812 827.870.747 327.120.820 9
    Table 6. Comparison of evaluation metrics of each method with scale factors of ×2, ×3, ×4
    比例方法参数量/MSet5Set14BSD100Urban100
    PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
    Bicubic-33.690.928 430.340.867 529.570.843 426.880.843 8
    SRCNN0.0236.660.954 232.420.906 331.360.887 929.500.894 6
    VDSR0.6737.530.958 733.030.912 431.900.896 030.760.914 0
    DRRN0.3037.740.959 133.230.913 632.050.897 331.230.918 8
    A2F-L1.3638.090.960 733.780.919 232.230.900 232.460.931 3
    JTF-SISR17.838.240.961 333.830.919 632.340.901 732.810.934 9
    ×2EDSR40.738.110.960 233.920.919 532.320.901 332.930.935 1
    RDN22.138.240.961 434.010.921 232.340.901 732.890.935 3
    E-GSCN15.738.210.961 233.950.920 332.350.901 932.940.935 7
    DRLN32.438.270.961 634.280.923 132.440.902 833.370.939 0
    SwinIR11.838.350.962 034.140.922 732.440.903 033.400.939 3
    MAAN9.6138.370.962 634.260.923 832.490.903 233.430.940 1
    Bicubic-30.410.865 527.640.772 227.210.734 424.460.741 1
    SRCNN0.0232.750.909 029.280.820 928.410.786 326.240.798 9
    VDSR0.6733.660.921 329.770.831 428.820.797 627.140.827 9
    DRRN0.3034.030.924 429.960.834 928.950.800 427.530.837 8
    A2F-L1.3734.540.928 330.410.843 629.140.806 228.400.857 4
    JTF-SISR18.134.720.929 530.570.846 529.250.809 028.740.864 5
    Table 6. Comparison of evaluation metrics of each method with scale factors of ×2, ×3, ×4
    方法PSNRSSIM

    时间

    /s

    参数量/M计算量/G
    EDSR27.710.742 00.33843.1516
    RDN27.720.741 90.22422.6209
    E-GSCN27.740.742 40.13615.9144
    DRLN27.830.744 40.27134.0292
    SwinIR27.830.745 90.11211.9128
    MAAN27.870.747 30.0859.7279
    Table 7. Complexity comparison of the 3 methods
    Ying ZHOU, Shenghu PEI, Haiyong CHEN, Shibo XU. Image super-resolution network based on multi-scale adaptive attention[J]. Optics and Precision Engineering, 2024, 32(6): 843
    Download Citation