Chinese Journal of Liquid Crystals and Displays, Volume. 38, Issue 9, 1234(2023)

Lightweight salt-and-pepper denoising combining noise mask training and nearest searching mechanism

Cheng-qiang HUANG* and Xing JIN
Author Affiliations
  • School of Physics and Electronics Sciences,Zunyi Normal University,Zunyi 563006,China
  • show less
    Figures & Tables(18)
    CNN for salt-and-pepper denoising
    Schematic diagram of image block interception method
    “Noise image”-“noise mask” pairs
    Construction method for noise “noise image”-“noise mask” pairs dataset
    Lightweight CNN
    Depth separable convolution
    Function of each convolution layer
    Function of each layer in depth separable convolution
    Structure chart of CMOS image sensor
    Example of the nearest searching mechanism
    Training curves.(a)Loss curve;(b)Accuracy curve.
    Comparison of noise mask images
    Comparison of visual effect for image processed by different denoising method
    • Table 1. Comparison of misjudging rate

      View table
      View in Article

      Table 1. Comparison of misjudging rate

      噪声图像极点标记8极值图像块标记7均值标记9CNN标记
      kodim20_0.11.154 2140.075 5301.154 2140.042 928
      kodim24_0.30.118 9340.000 3220.118 9250.010 885
      kodim18_0.50.000 9410.000 9410.002 9090.002 655
      kodim08_0.70.009 7870.049 6460.009 2350.008 552
      kodim13_0.90.003 0550.283 3110.000 5620.001 986
      均值0.257 3860.081 9500.257 1690.013 401
    • Table 2. Comparison of denoising performance with the same noise marking

      View table
      View in Article

      Table 2. Comparison of denoising performance with the same noise marking

      去噪方法指标kodim20kodim24kodim18平均值
      中值滤波PSNR38.8324.0415.7126.19
      MSE0.383.6813.915.99
      三权重因子PSNR32.9927.4227.3829.26
      MSE0.582.714.162.48
      自适应概率滤波PSNR34.8028.2524.6729.24
      MSE0.542.655.152.78
      改善的中值滤波PSNR23.7123.4126.6624.59
      MSE1.533.404.503.14
      均值标定重复滤波PSNR35.1725.6426.7129.17
      MSE0.532.904.422.62
      最近邻搜索PSNR35.1229.3527.5030.65
      MSE0.532.464.112.37
    • Table 3. Comparison of denoising performance of various methods

      View table
      View in Article

      Table 3. Comparison of denoising performance of various methods

      算法指标kodim05kodim24kodim23均值
      中值滤波5PSNR24.1914.648.0715.63
      MSE7.5120.0269.6632.40
      三权重因子8PSNR29.8120.3527.8326.00
      MSE2.046.513.303.95
      自适应概率滤波7PSNR29.2424.2511.1621.55
      MSE2.165.2435.5614.32
      改善的中值滤波6PSNR28.9920.2627.325.52
      MSE2.246.883.604.24
      均值标定重复滤波9PSNR29.6718.327.2325.07
      MSE1.998.853.744.86
      Liang CNN17PSNR21.5121.6524.7122.62
      MSE14.1613.569.1012.27
      Xing CNN16PSNR27.4325.2928.6927.14
      MSE6.857.775.416.68
      本文方法PSNR29.8926.3328.9728.40
      MSE2.044.443.203.23
    • Table 4. Comparison of denoising performance on BSD300 dataset

      View table
      View in Article

      Table 4. Comparison of denoising performance on BSD300 dataset

      去噪方法PSNRMSE
      中值滤波517.2330.21
      三权重因子829.664.44
      自适应概率滤波724.6615.24
      改善的中值滤波628.444.86
      均值标定重复滤波929.624.57
      Liang CNN1722.5114.03
      Xing CNN1627.917.56
      本文方法30.414.14
    • Table 5. Comparison of network complexity

      View table
      View in Article

      Table 5. Comparison of network complexity

      网络MFLOPS参数数量

      .h5文件

      大小/kB

      Liang CNN1717 372.033 572 99343 712
      Xing CNN162 717.66558 9776 840
      本文CNN168.2538 401723
    Tools

    Get Citation

    Copy Citation Text

    Cheng-qiang HUANG, Xing JIN. Lightweight salt-and-pepper denoising combining noise mask training and nearest searching mechanism[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(9): 1234

    Download Citation

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

    Category: Research Articles

    Received: Oct. 26, 2022

    Accepted: --

    Published Online: Sep. 19, 2023

    The Author Email: Cheng-qiang HUANG (18616836345@163.com)

    DOI:10.37188/CJLCD.2022-0356

    Topics