• Journal of Atmospheric and Environmental Optics
  • Vol. 19, Issue 3, 381 (2024)
WU Xiaohua1, LI Zenglu2,3, XU Zhanghua4, and ZHOU Jingchun5,*
Author Affiliations
  • 1School of Art & Design, Sanming University, Sanming 365004, China
  • 2Network Technology Center, Sanming University, Sanming 365004, China
  • 3Fujian Provincial Key Laboratory of Resources and Environment Monitoring & Sustainable Management and Utilization,Sanming University, Sanming 365004, China
  • 4Academy of Geography and Ecological Environment, Fuzhou University, Fuzhou 350108, China
  • 5Information Science and Technology College, Dalian Maritime University, Dalian 116026, China
  • show less
    DOI: 10.3969/j.issn.1673-6141.2024.03.010 Cite this Article
    Xiaohua WU, Zenglu LI, Zhanghua XU, Jingchun ZHOU. Lightweight underwater image enhancement network based on cross-scale deep distillation feature perception[J]. Journal of Atmospheric and Environmental Optics, 2024, 19(3): 381 Copy Citation Text show less
    Lightweight underwater image enhancement network structure diagram based on CSDDFP module
    Fig. 1. Lightweight underwater image enhancement network structure diagram based on CSDDFP module
    Comparisons of parameter size and objective performance on LSUI400 dataset
    Fig. 2. Comparisons of parameter size and objective performance on LSUI400 dataset
    Visual comparisons of the mainstream methods on four benchmark datasets. (a) Input image; (b) UGAN; (c) WaterNet; (d) FUnIE_GAN; (e) UT-UIE; (f) ours; (g) ground truth
    Fig. 3. Visual comparisons of the mainstream methods on four benchmark datasets. (a) Input image; (b) UGAN; (c) WaterNet; (d) FUnIE_GAN; (e) UT-UIE; (f) ours; (g) ground truth
    Visual comparisons of the mainstream methods under insufficient illumination in deep water. (a) Input image; (b) UGAN; (c) WaterNet; (d) FUnIE_GAN; (e) UT-UIE; (f) ours; (g) ground truth
    Fig. 4. Visual comparisons of the mainstream methods under insufficient illumination in deep water. (a) Input image; (b) UGAN; (c) WaterNet; (d) FUnIE_GAN; (e) UT-UIE; (f) ours; (g) ground truth
    Visual comparisons of the mainstream methods on texture details. (a) Input image; (b) UGAN; (c) WaterNet; (d) FUnIE_GAN; (e) UT-UIE; (f) ours; (g) ground truth
    Fig. 5. Visual comparisons of the mainstream methods on texture details. (a) Input image; (b) UGAN; (c) WaterNet; (d) FUnIE_GAN; (e) UT-UIE; (f) ours; (g) ground truth
    MethodsLSUI400EUVP_Test515UIEB100OceanEx
    PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
    UWCNN17.3660.72517.7250.70414.1550.68615.9600.724
    Cycle-GAN18.3200.74917.9630.70917.7140.75821.0070.828
    SGUIE-Net19.9100.81919.1870.76021.1780.87218.6770.834
    FUnIE-GAN23.2720.81824.0770.79419.6140.81320.4480.855
    UGAN25.1170.84623.6360.80521.3680.82522.4360.822
    UT-UIE24.3490.82925.2120.81320.9160.76421.2690.822
    Ours26.6720.87225.3940.83322.7900.87722.0030.876
    Table 1. Objective comparisons of different underwater image enhancement methods
    KP/MCSDDFPSGFM[15]PPM[23]
    PSNRSSIMPSNRSSIMPSNRSSIM
    01.3425.6890.86425.6890.86425.6890.864
    12.9926.4390.87026.0120.86625.9020.865
    24.6426.6720.87226.1700.86726.0840.866
    Table 2. Ablation studies of the cross-scale deep distillation feature perception module
    Xiaohua WU, Zenglu LI, Zhanghua XU, Jingchun ZHOU. Lightweight underwater image enhancement network based on cross-scale deep distillation feature perception[J]. Journal of Atmospheric and Environmental Optics, 2024, 19(3): 381
    Download Citation