• Laser & Optoelectronics Progress
  • Vol. 57, Issue 14, 141027 (2020)
Bingyuan Wang1、2, Fang Zheng2、*, Jian Jiang2, and Bo Yang2
Author Affiliations
  • 1Ground Support Equipment Research Base of Civil Aviation University of China, Tianjin 300300, China
  • 2School of Electronic Information and Automation, Civil Aviation University of China, Tianjin 300300, China
  • show less
    DOI: 10.3788/LOP57.141027 Cite this Article Set citation alerts
    Bingyuan Wang, Fang Zheng, Jian Jiang, Bo Yang. Method for Removal of Rain and Fog in Single Image[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141027 Copy Citation Text show less
    Mapping of manifold learning
    Fig. 1. Mapping of manifold learning
    Mapping and embedding of LLE algorithm
    Fig. 2. Mapping and embedding of LLE algorithm
    Dehazing map and its color histogram. (a) Foggy image; (b) manifold particle filtering for fog removal; (c) color histogram of Fig. 3(a); (d) color histogram of Fig. 3(b)
    Fig. 3. Dehazing map and its color histogram. (a) Foggy image; (b) manifold particle filtering for fog removal; (c) color histogram of Fig. 3(a); (d) color histogram of Fig. 3(b)
    Architecture of proposed method in this paper
    Fig. 4. Architecture of proposed method in this paper
    Original images and test results of each algorithm for synthetic uniform fog and fog in natural scene. (a) Input hazy images; (b) soft matting; (c) GF algorithm; (d) method proposed by He et al.; (e) Retinex algorithm; (f) HF; (g) NBPC+PA; (h) proposed method
    Fig. 5. Original images and test results of each algorithm for synthetic uniform fog and fog in natural scene. (a) Input hazy images; (b) soft matting; (c) GF algorithm; (d) method proposed by He et al.; (e) Retinex algorithm; (f) HF; (g) NBPC+PA; (h) proposed method
    Original images and test results of each algorithm for rain and fog removal. (a) Rain image; (b) Gaussian curvature algorithm; (c) DRN; (d) proposed method
    Fig. 6. Original images and test results of each algorithm for rain and fog removal. (a) Rain image; (b) Gaussian curvature algorithm; (c) DRN; (d) proposed method
    LayerFeature
    0Detail layer
    1Conv(1×1, 32); stride:1; ReLU
    2Conv(1×1, 32); stride:1; ReLU
    3Conv(1×1, 32); stride: 1; ReLU
    Residual4Concatenate (2,3)
    5Conv(1×1, 32); stride:1; ReLU
    6Conv (1×1, 32); stride:1; ReLU
    7Concatenate (4,5)
    Input gateConv (3×3, 32); stride:1; sigmoid
    LSTMForget gateConv (3×3, 32); stride:1; sigmoid
    Cell stateConv (3×3, 32); stride:1; sigmoid
    Output gateConv (3×3, 32); stride:1; sigmoid
    Table 1. Structural parameter settings of recursive network
    LayerFeature
    Layer 0Output image
    Layer 1Conv (5×5, 8); stride:1; ReLU
    Layer 2Conv (5×5, 16); stride:1; ReLU
    Layer 3Conv (5×5, 32); stride:1; ReLU
    Layer 4Conv (5×5, 64); stride:1; ReLU
    Layer 5Conv (5×5, 128); stride:1; ReLU
    Layer 6Conv (5×5, 128); stride:1; ReLU
    Layer 7Conv (5×5, 64); stride:4; ReLU
    Layer 8Conv (5×5, 64); stride:4; ReLU
    Layer 9Conv (5×5, 32); stride:4; ReLU
    Table 2. Discriminator structural parameter setting
    SettingNo.SoftmattingGF algorithmMethodproposedby He et al.RetinexalgorithmHFNBPC+PAMPF (manifoldparticlefiltering)
    MSD /1050.8320.8941.3100.9950.8571.2122.011
    1PSNR62.45461.27260.83260.58262.65861.53662.720
    SSIM0.4180.4030.4160.2510.6900.5340.746
    MSD /1050.9540.9801.5381.0580.9760.7791.310
    2PSNR59.08160.71859.93058.37358.54560.68161.346
    SSIM0.4500.4840.3890.3730.4730.5150.815
    MSD /1052.5903.4164.2353.2202.3191.9605.099
    3PSNR63.07458.25861.47454.30760.50563.59165.186
    SSIM0.8128.5160.5390.3250.7560.8180.871
    MSD /1052.1222.7473.4621.6562.7031.8864.990
    4PSNR61.74756.11460.24955.53261.31963.49861.275
    SSIM0.5890.4740.4830.3510.5950.6840.743
    MSD /1052.4943.2833.7003.0761.9731.7834.495
    5PSNR57.72156.29257.98754.65461.39660.75660.127
    SSIM0.5300.4600.5380.3330.6990.8090.828
    MSD /1053.6345.3885.7454.6884.0523.4787.636
    6PSNR65.72957.37258.64454.45660.79465.20767.130
    SSIM0.7260.5880.6160.3340.7010.8200.905
    Table 3. Data obtained by objective evaluation methods for testing images in Fig. 5
    Scene No.FSIMSSIMPSNR
    GCDRNIA-GANGCDRNIA-GANGCDRNIA-GAN
    10.7410.7780.9670.7400.9060.93764.03837.48271.495
    20.8120.7940.8420.9030.8960.95667.36870.04870.544
    30.7030.8330.8950.9000.9120.94766.61170.68174.159
    40.8040.8640.8470.9070.8870.94670.81169.24369.879
    50.7180.8730.8930.3540.9150.93968.33771.08672.815
    注:GC: Gaussian curvature; IA-GAN: improved attentive generative adversarial network
    Table 4. Data obtained by objective evaluation methods for natural rain and fog scene for experiments (Fig. 6)
    Bingyuan Wang, Fang Zheng, Jian Jiang, Bo Yang. Method for Removal of Rain and Fog in Single Image[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141027
    Download Citation