Laser & Optoelectronics Progress, Volume. 62, Issue 16, 1615001(2025)

Super-Resolution Reconstruction and Denoising Tasks for Public Safety Scene Images Using the EnSwinIR Model

Qixiang Meng1, Fanliang Bu1、*, and Qiqi Kou2
Author Affiliations
  • 1School of Information Network Security, People's Public Security University of China, Beijing 100038, China
  • 2School of Computer Science and Technology, China University of Mining and Technology, Xuzhou 221116, Jiangsu , China
  • show less
    Figures & Tables(13)
    Schematic diagrams of SwinIR super-resolution network structure. (a) The overall flowchart of SwinIR; (b) residual SwinIR Transformer block; (c) Swin Transformer layer
    Schematic diagrams of SwinIR super-resolution network structure. (a) The overall flowchart of EnSwinIR; (b) hybrid residual attention block; (c) shift convolution; (d) group multi-scale self-attention
    2× super-resolution reconstruction training change curves
    4× super-resolution reconstruction training change curves
    Visual comparison of 2× super-resolution reconstruction models
    Visual comparison of 4× super-resolution reconstruction models
    Comparison of denoising effects of different models on color images at a high noise level (σ=50)
    • Table 1. Hardware configuration

      View table

      Table 1. Hardware configuration

      HardwareType
      GPUA800-80 GB
      CPU14 vCPU Intel(R) Xeon(R) Gold 6348
      Memory100 GB
      StorageSystem disk: 30 GB; data disk: 300 GB
    • Table 2. Software configuration

      View table

      Table 2. Software configuration

      SoftwareType
      Operating systemUbuntu 22.04
      Deep learning frameworkPyTorch 2.1.2
      Deep learning libraryCUDA 11.8
      Image processing libraryOpenCV 4.5.3
      Python versionPython 3.9.7
    • Table 3. Super-resolution reconstruction model performance indicators

      View table

      Table 3. Super-resolution reconstruction model performance indicators

      ScaleModelParams /103FLOPS /109PSNR/SSIM
      MallStreetGovernmentAverage
      ×2SRCNN157.13.7423.6/0.60322.8/0.58923.1/0.61423.17/0.602
      EDSR24073023129.7/0.85128.5/0.85328.9/0.84829.03/0.851
      HAN1363608194.5933.8/0.87632.4/0.87132.7/0.87432.97/0.874
      SwinIR1411630194.634.1/0.89333.6/0.88734.2/0.89533.97/0.892
      HAT172062410234.4/0.89734.2/0.89334.3/0.88934.30/0.893
      MAN2087318334.7/0.90534.3/0.90134.6/0.90734.53/0.904
      DAT181106520735.2/0.91135.4/0.90835.1/0.90235.23/0.907
      RGT191312024335.6/0.91635.2/0.90935.1/0.90435.30/0.910
      EnSwinIR8429134.8136.1/0.92135.6/0.92335.9/0.91835.87/0.921
      ×4SRCNN157.33.7620.4/0.51421.1/0.50921.5/0.51721.00/0.513
      EDSR24309023323.6/0.77723.8/0.77123.9/0.77823.77/0.775
      HAN136419919625.4/0.78125.1/0.77625.7/0.78125.40/0.779
      SwinIR141166721128.3/0.79928.8/0.80129.2/0.80728.77/0.802
      HAT172077110428.6/0.80528.5/0.80729.0/0.80428.70/0.805
      MAN2087558228.8/0.81128.4/0.80828.6/0.81528.60/0.811
      DAT181121421629.3/0.82429.5/0.82929.7/0.83129.50/0.828
      RGT191332025130.4/0.83330.2/0.83130.1/0.83530.23/0.833
      EnSwinIR8487135.7731.7/0.85131.8/0.85431.5/0.85231.67/0.852
    • Table 4. Ablation experimental results of local feature extraction module

      View table

      Table 4. Ablation experimental results of local feature extraction module

      ScaleParams /103FLOPS /109Shift-ConvResidualPSNR /dBSSIM
      LeftRightUpperDown
      ×23569.0858.0528.630.784
      8416.37132.9135.110.906
      8416.33132.8935.070.908
      8428.96133.6735.620.911
      8429.07134.8135.870.921
      ×43627.4359.0124.410.723
      8462.28134.0831.360.844
      8462.41134.1531.380.842
      8487.28135.4331.550.847
      8487.32135.7731.670.852
    • Table 5. Ablation experimental results of GMSA

      View table

      Table 5. Ablation experimental results of GMSA

      ScaleParams /103FLOPS /109Multi-scale window sizeResidualPSNR /dBSSIM
      Window_oneWindow_twoWindow_three
      ×24884.5277.1626.710.761
      8416.88133.484×44×44×435.440.913
      8430.67135.048×88×88×835.730.918
      8626.93146.8812×1212×1212×1235.960.925
      8428.15134.554×48×812×1235.490.915
      8429.07134.814×48×812×1235.870.921
      ×44942.8778.1222.450.727
      8445.12134.934×44×44×431.070.843
      8489.14136.118×88×88×831.750.854
      8693.18147.4912×1212×1212×1231.980.859
      8486.89135.124×48×812×1231.030.841
      8487.32135.774×48×812×1231.670.852
    • Table 6. Quantitative average PSNR of color image denoising methods

      View table

      Table 6. Quantitative average PSNR of color image denoising methods

      SceneσDncNN24FFDNet25Noise2Noise11SwinIR14LAN26Proposed method
      Mall1524.1325.5126.3729.1631.2232.96
      2522.0523.6624.7927.3529.4830.17
      5020.5922.4222.9226.2428.6429.44
      Government1525.0425.6326.2629.3831.5532.78
      2522.1823.1124.8127.2429.8930.89
      5020.8622.0722.9326.0929.1230.26
      Street1523.8225.1226.4228.4331.3832.44
      2522.7123.8224.3427.5329.4930.77
      5020.8922.3622.5725.9228.7729.97
      Average1524.3325.4226.3528.9631.3832.73
      2522.3123.5324.6527.3729.6230.61
      5020.7822.2822.8126.0828.8429.89
    Tools

    Get Citation

    Copy Citation Text

    Qixiang Meng, Fanliang Bu, Qiqi Kou. Super-Resolution Reconstruction and Denoising Tasks for Public Safety Scene Images Using the EnSwinIR Model[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1615001

    Download Citation

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

    Category: Machine Vision

    Received: Dec. 5, 2024

    Accepted: Feb. 7, 2025

    Published Online: Aug. 18, 2025

    The Author Email: Fanliang Bu (20051257@ppsuc.edu.cn)

    DOI:10.3788/LOP242377

    CSTR:32186.14.LOP242377

    Topics