• Optics and Precision Engineering
  • Vol. 30, Issue 13, 1606 (2022)
Beibei SONG1,*, Suina MA1, Fan HE1, and Wenfang SUN2
Author Affiliations
  • 1School of Information Engineering, Chang'an University, Xi'an70064, China
  • 2School of Aerospace Science and Technology, Xidian University, Xi'an71016, China
  • show less
    DOI: 10.37188/OPE.2021.0433 Cite this Article
    Beibei SONG, Suina MA, Fan HE, Wenfang SUN. Hyperspectral reconstruction from RGB images based on Res2-Unet deep learning network[J]. Optics and Precision Engineering, 2022, 30(13): 1606 Copy Citation Text show less
    Res2Net module
    Fig. 1. Res2Net module
    Network architecture of Res2-Unet
    Fig. 2. Network architecture of Res2-Unet
    Res2Net-SE module and SE block
    Fig. 3. Res2Net-SE module and SE block
    SAM comparison of ARAD_HS_0451 data on Clean track
    Fig. 4. SAM comparison of ARAD_HS_0451 data on Clean track
    SAM comparison of ARAD_HS_0463 data on Clean track
    Fig. 5. SAM comparison of ARAD_HS_0463 data on Clean track
    SAM comparison of ARAD_HS_0451 data on Real World track
    Fig. 6. SAM comparison of ARAD_HS_0451 data on Real World track
    SAM comparison of ARAD_HS_0463 data on Real World track
    Fig. 7. SAM comparison of ARAD_HS_0463 data on Real World track
    Spectral curve comparison of ARAD_HS_0451 on Clean track
    Fig. 8. Spectral curve comparison of ARAD_HS_0451 on Clean track
    Spectral curve comparison of ARAD_HS_0463 on Clean track
    Fig. 9. Spectral curve comparison of ARAD_HS_0463 on Clean track
    Spectral curve comparison of ARAD_HS_0451 on Real World track
    Fig. 10. Spectral curve comparison of ARAD_HS_0451 on Real World track
    Spectral curve comparison of ARAD_HS_0463 on Real World track
    Fig. 11. Spectral curve comparison of ARAD_HS_0463 on Real World track
    MethodMRAERMSEPSNRMSAM

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation
    AWAN0.034 30.017 20.011 80.007 540.260 15.814 42.349 11.046 1
    HRNet0.039 60.017 60.014 10.009 038.610 95.747 02.680 01.061 8
    Res2-Unet0.034 00.012 20.011 70.008 240.348 05.759 82.267 60.787 4
    Table 1. Comparison of test results for Clean track
    MethodMRAERMSEPSNRMSAM

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation
    AWAN0.066 10.019 40.017 80.009 736.119 14.614 93.336 61.181 1
    HRNet0.071 40.020 20.018 10.009 535.867 24.405 53.636 81.202 9
    Res2-Unet0.066 10.019 00.016 10.008 236.840 04.233 93.240 11.005 8
    Table 2. Comparison of test results for Real World track
    IndexAWANHRNetRes2-Unet
    MRAERMSEPSNRMSAMMRAERMSEPSNRMSAMMRAERMSEPSNRMSAM
    MRAE1.000 00.759 0-0.751 10.964 51.000 00.800 1-0.739 20.977 91.000 00.644 3-0.541 30.958 3
    RMSE0.759 01.000 0-0.958 00.743 30.800 11.000 0-0.947 90.752 60.644 31.000 0-0.957 00.653 5
    PSNR-0.751 1-0.958 01.000 0-0.758 8-0.739 2-0.947 91.000 0-0.728 5-0.541 3-0.957 01.000 0-0.518 0
    MSAM0.964 50.743 3-0.758 81.000 00.977 90.752 6-0.728 51.000 00.958 30.653 5-0.518 01.000 0
    Table 3. Correlation coefficients of test results on Clean track
    IndexAWANHRNetRes2-Unet
    MRAERMSEPSNRMSAMMRAERMSEPSNRMSAMMRAERMSEPSNRMSAM
    MRAE1.000 00.373 7-0.265 50.943 91.000 00.371 7-0.251 50.951 41.000 00.184 2-0.069 60.938 9
    RMSE0.373 71.000 0-0.983 10.521 40.371 71.000 0-0.986 30.488 40.184 21.000 0-0.987 10.377 7
    PSNR-0.265 5-0.983 11.000 0-0.438 7-0.251 5-0.986 31.000 0-0.396 5-0.069 6-0.987 11.000 0-0.286 2
    MSAM0.943 90.521 4-0.438 71.000 00.951 40.488 4-0.396 51.000 00.938 90.377 7-0.286 21.000 0
    Table 4. Correlation coefficients of test results on Real World track
    MethodMRAERMSEPSNRSAM

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation

    Mean

    value

    Standard deviation
    Conv3×30.071 30.022 30.019 90.011 035.143 64.692 74.093 81.168 4
    Res2Net0.043 90.021 80.015 70.011 338.041 56.152 42.972 81.272 3
    Res2Net-SE0.034 00.012 20.011 70.008 240.348 05.759 82.267 60.787 4
    Table 5. Comparison of test results of network ablation on Clean track
    MethodMRAERMSEPSNRSAM
    Mean valueStandard deviationMean valueStandard deviationMean valueStandard deviationMean valueStandard deviation
    Conv3×30.086 40.023 40.020 00.010 234.967 34.329 24.225 71.192 7
    Res2Net0.068 30.017 90.018 30.008 935.685 04.292 03.476 41.010 0
    Res2Net-SE0.066 10.019 00.016 10.008 236.840 04.233 93.240 11.005 8
    Table 6. Comparison of test results of network ablation on Real World track
    Beibei SONG, Suina MA, Fan HE, Wenfang SUN. Hyperspectral reconstruction from RGB images based on Res2-Unet deep learning network[J]. Optics and Precision Engineering, 2022, 30(13): 1606
    Download Citation