Opto-Electronic Engineering, Volume. 50, Issue 2, 220185(2023)

Light field image super-resolution network based on angular difference enhancement

Tianqi Lv... Yingchun Wu* and Xianling Zhao |Show fewer author(s)
Author Affiliations
  • School of Electronic and Information Engineering, Taiyuan University of Science and Technology, Taiyuan, Shanxi 030024, China
  • show less
    Figures & Tables(17)
    4D light field acquisition and rearrangement. (a) Biplanar representation model of light field; (b) Subaperture image array; (c) Macropixel array
    Overall network structure diagram
    Multi-branch residual block
    ADA module details. (a) ADA module data processing process; (b) Feature collection; (c) Offset acquisition in feature collection; (d) Update the central view; (e) Feature distribution; (f) Offset acquisition in feature distribution
    EADA module details. (a) EADA feature collection details; (b) EADA feature distribution details
    RFD module simplification. (a) RFD module details; (b) SRFD module details
    Visual contrast of the "Origami" scene with 2× SR
    Visual contrast of the "Herbs" scene with 2× SR
    Visual contrast of the "Bee" scene with 4× SR
    Visual contrast of the "Lego Knights" scene with 4× SR
    • Table 1. Five public light field datasets used in our experiment

      View table
      View in Article

      Table 1. Five public light field datasets used in our experiment

      数据集EPFL[13]HCInew[14]HCIold[15]INRIA[16]STFgantry[17]总共
      #训练702010359144
      #测试10425223
    • Table 2. PSNR/SSIM values achieved by different shallow feature extraction modules for 4×SR

      View table
      View in Article

      Table 2. PSNR/SSIM values achieved by different shallow feature extraction modules for 4×SR

      模型EPFL[13]HCInew[14]HCIold[15]INRIA[16]STFgantry[17]
      无残差25.26/0.832427.71/0.851732.58/0.934426.95/0.886726.09/0.8452
      MBR28.81/0.919031.30/0.920637.39/0.972530.81/0.951331.29/0.9511
    • Table 3. PSNR/SSIM values achieved by different deep feature extraction modules for 4× SR

      View table
      View in Article

      Table 3. PSNR/SSIM values achieved by different deep feature extraction modules for 4× SR

      模型EPFL[13]HCInew[14]HCIold[15]INRIA[16]STFgantry[17]
      ADA模块28.77/0.917231.26/0.919837.41/0.972330.80/0.950731.17/0.9497
      单支路EADA模块28.78/0.916631.21/0.918637.31/0.971730.85/0.950231.03/0.9474
      EADA模块28.81/0.919031.30/0.920637.39/0.972530.81/0.951331.29/0.9511
    • Table 4. PSNR/SSIM values achieved by different feature fusion modules for 4×SR

      View table
      View in Article

      Table 4. PSNR/SSIM values achieved by different feature fusion modules for 4×SR

      模型EPFL[13]HCInew[14]HCIold[15]INRIA[16]STFgantry[17]
      RFD模块29.01/0.918331.32/0.919837.39/0.971831.08/0.950931.14/0.9499
      SRFD模块28.81/0.919031.30/0.920637.39/0.972530.81/0.951331.29/0.9511
    • Table 5. PSNR/SSIM values achieved by different methods for 2× SR

      View table
      View in Article

      Table 5. PSNR/SSIM values achieved by different methods for 2× SR

      超分方法EPFL[13]HCInew[14]HCIold[15]INRIA[16]STFgantry[17]Average
      EDSR[20]33.09/0.963134.83/0.959441.01/0.987534.97/0.976536.29/0.981936.04/0.9728
      RCAN[21]33.16/0.963534.98/0.960241.05/0.987535.01/0.976936.33/0.982536.11/0.9741
      ResLF[8]32.75/0.967236.07/0.971542.61/0.992234.57/0.978436.89/0.987336.58/0.9793
      LFSSR[7]33.69/0.974836.86/0.975343.75/0.993935.27/0.983438.07/0.990237.53/0.9835
      LF-InterNet[4]34.14/0.976137.28/0.976944.45/0.994535.80/0.984638.72/0.991638.08/0.9847
      LF-DFnet[5]34.44/0.976637.44/0.978644.23/0.994336.36/0.9841 39.61/0.993538.42/0.9854
      本文方法34.58/0.977237.92/0.979644.84/0.994836.59/0.985440.11/0.993938.81/0.9862
    • Table 6. PSNR/SSIM values achieved by different methods for 4× SR

      View table
      View in Article

      Table 6. PSNR/SSIM values achieved by different methods for 4× SR

      超分方法EPFL[13]HCInew[14]HCIold[15]INRIA[16]STFgantry[17]Average
      EDSR[20]27.84/0.885829.60/0.887435.18/0.953829.66/0.925928.70/0.907530.20/0.9121
      RCAN[21]27.88/0.886329.63/0.888035.20/0.954029.76/0.927328.90/0.911030.27/0.9133
      ResLF[8]27.46/0.889929.92/0.901136.12/0.965129.64/0.933928.99/0.921430.43/0.9223
      LFSSR[7]28.27/0.908030.72/0.912436.70/0.969030.31/0.944630.15/0.938531.23/0.9345
      LF-InterNet[4]28.67/0.914330.98/0.916537.11/0.971530.64/0.948630.53/0.942631.59/0.9387
      LF-DFnet[5]28.77/0.916531.23/0.919637.32/0.971830.83/0.950331.15/0.949431.86/0.9415
      本文方法28.81/0.919031.30/0.920637.39/0.972530.81/0.951331.29/0.951131.92/0.9429
    • Table 7. Comparisons of the number of parameters and FLOPs by different methods for 2× SR and 4× SR

      View table
      View in Article

      Table 7. Comparisons of the number of parameters and FLOPs by different methods for 2× SR and 4× SR

      MethodParameters/MFLOPs/G
      EDSR[20]38.62/38.8939.56×25/40.66×25
      RCAN[21]15.31/15.3615.59×25/15.65×25
      ResLF[8]6.35/6.7937.06/39.70
      LFSSR[7]0.81/1.6125.70/128.44
      LF-InterNet[4]4.80/5.2347.46/50.10
      LF-DFnet[5]3.94/3.9957.22/57.31
      本文方法12.74/12.80238.92/240.51
    Tools

    Get Citation

    Copy Citation Text

    Tianqi Lv, Yingchun Wu, Xianling Zhao. Light field image super-resolution network based on angular difference enhancement[J]. Opto-Electronic Engineering, 2023, 50(2): 220185

    Download Citation

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

    Category: Article

    Received: Jul. 28, 2022

    Accepted: Oct. 21, 2022

    Published Online: Apr. 13, 2023

    The Author Email: Wu Yingchun (yingchunwu3030@foxmail.com)

    DOI:10.12086/oee.2023.220185

    Topics