• Optics and Precision Engineering
  • Vol. 32, Issue 12, 1954 (2024)
Mingzhu SHI1,2,*, Bin ZAO1,2, Yuhao SU1,2, Xinhui LIN1,2..., Siqi KONG1,2 and Muxian TAN1,2|Show fewer author(s)
Author Affiliations
  • 1College of Electronic and Communication Engineering, Tianjin Normal University, Tianjin300387, China
  • 2Tianjin Key Laboratory of Wireless Mobile Communications and Power Transmission, Tianjin Normal University, Tianjin300387, China
  • show less
    DOI: 10.37188/OPE.20243212.1954 Cite this Article
    Mingzhu SHI, Bin ZAO, Yuhao SU, Xinhui LIN, Siqi KONG, Muxian TAN. Dual attention refinement single image desnowing[J]. Optics and Precision Engineering, 2024, 32(12): 1954 Copy Citation Text show less
    Structure of dual Attention refinement desnowing network
    Fig. 1. Structure of dual Attention refinement desnowing network
    Channel attention and pixel attention
    Fig. 2. Channel attention and pixel attention
    Gated-Dconv feed-forward (GDFN)
    Fig. 3. Gated-Dconv feed-forward (GDFN)
    Results of snow removal by different methods on the CSD dataset
    Fig. 4. Results of snow removal by different methods on the CSD dataset
    Visualization comparison of different snow removal methods for texture details
    Fig. 5. Visualization comparison of different snow removal methods for texture details
    Visual comparison with other methods on real snow images
    Fig. 6. Visual comparison with other methods on real snow images
    Results of ablation experiments of branch network and refinement processing module for snow removal
    Fig. 7. Results of ablation experiments of branch network and refinement processing module for snow removal
    方法CSD(2 000)SRRS(2 000)Snow100K(2 000)#Param# GMacs
    PSNR/dBSSIMPSNR/dBSSIMPSNR/dBSSIM
    Zheng220.210.7919.120.8023.860.82--
    DesnowNet520.130.8120.380.8430.500.9415.6 M1.7 KG
    CycleGAN2120.980.8020.210.7426.810.897.84 M42.38 G
    RESCAN2222.110.8122.790.86----
    All in One2326.310.8724.980.8826.070.8844 M12.26 G
    JSTASR627.960.8825.820.8923.120.8665 M-
    HDCWNet729.060.9127.780.9231.540.956.99 M9.78 G
    TransWeather2431.760.9328.290.9231.820.9321.9 M5.64 G
    Ours32.160.9630.080.9532.030.955.48 M6.12 G
    Table 1. Results of different desnowing algorithms on three datesets
    指标PSNR/dBSSIM
    分支网络串联结构29.830.94
    维度不拆分并行化策略31.940.95
    维度拆分并行化策略32.160.96
    Table 2. Comparison results of different construction strategies
    指标像素注意力空间注意力通道注意力
    PSNR/dB32.0531.9432.16
    SSIM0.960.960.96
    Table 3. Comparison results of different attention strategies
    纹理提取分支

    精细化

    处理模块

    U型特征提取分支网络

    指标

    (PSNR/SSIM)

    ××30.860.94
    ×31.430.95
    ×31.870.95
    32.160.96
    Table 4. Results of ablation experiments on branching networks and refinement of processing modules
    Mingzhu SHI, Bin ZAO, Yuhao SU, Xinhui LIN, Siqi KONG, Muxian TAN. Dual attention refinement single image desnowing[J]. Optics and Precision Engineering, 2024, 32(12): 1954
    Download Citation