Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 7, 929(2024)

Correction of vignetting images based on Retinex-Net network model

Dandan HUANG1,2, Fei WANG1,2, Zhi LIU1,2、*, Han GAO1,2, and Huiji WANG1,2
Author Affiliations
  • 1School of Electronic Information Engineering,Changchun University of Technology,Changchun 130022,China
  • 2Space Optoelectronics Technology National and Local Joint Engineering Research Center,Changchun University of Technology,Changchun 130022,China
  • show less
    Figures & Tables(11)
    Schematic diagram of gradual halo phenomenon
    Improved Retinex-Net network model
    Diagram of RDB structure
    Schematic diagram of residual block structure
    Examples of vignetting image dataset
    Histograms and function curves of the image
    Comparison of different correction algorithms for processing vignetting images
    • Table 1. Decomposition network

      View table
      View in Article

      Table 1. Decomposition network

      层次结构激活函数卷积核大小步长作用
      输入层输入渐晕图像和对照图像
      隐藏层卷积3×32特征提取
      卷积+ReLu激活3×32利用两个约束条件:反射不变约束、光照平滑约束得到IR
      卷积+ReLu激活3×32
      输出层卷积3×32从提取的多通道特征中投影出RI
      sigmoid约束变量,保证阈值为正
    • Table 2. Calibration network

      View table
      View in Article

      Table 2. Calibration network

      增强子网络
      层次结构激活函数卷积核大小步长作用
      输入层输入渐晕图像的光照分量图像/反射分量图像
      隐藏层卷积+ReLu激活3×32下采样:提高感受野,明确大尺度的光照分量情况
      卷积+ReLu激活3×32与上采样快镜像连接,重建局部光照分布情况
      插值+卷积+ReLu3×32镜像上采样块:重建局部光照分布,强制网络学习残差
      插值+卷积+ReLu3×32
      卷积1×12将多尺度拼接后的特征通道简化为c通道
      输出层卷积3×32重建光照图
      去噪子网络
      层次结构激活函数卷积核大小步长作用
      输入层输入渐晕图像的反射分量图像
      隐藏层卷积+ReLu激活3×31浅层特征提取
      RDB具有连续记忆机制,通过局部密集连接利用所有层提取到的特征
      卷积+批规范化+ReLu3×31深层特征提取,生成非线性特征映射,提升网络精度,加快网络收敛速度
      卷积+批规范化+ReLu卷积+批规范化+ReLu卷积+批规范化+ReLu3×33×33×3111
      输出层卷积+批规范化+ReLu3×31利用残差学习策略训练,输出去噪后的反射图
      卷积3×31
    • Table 3. Comparative experiment on ablation performance of improved Retinex-Net network model

      View table
      View in Article

      Table 3. Comparative experiment on ablation performance of improved Retinex-Net network model

      网络模型SSIMPSNRRMSE
      Retinex-Net0.53316.3400.636
      Retinex-Net+dilated Conv0.54616.9600.533
      Retinex-Net+RRDB0.79917.0220.624
      Retinex-Net+dilated Conv+RRDB0.82617.0670.541
    • Table 4. Evaluation indicators of corrected images obtained by different correction methods

      View table
      View in Article

      Table 4. Evaluation indicators of corrected images obtained by different correction methods

      方法SSIMPSNRRMSE
      文献[170.54314.7620.599
      文献[180.48616.4390.620
      文献[190.62716.9600.715
      本文方法0.82617.0670.541
    Tools

    Get Citation

    Copy Citation Text

    Dandan HUANG, Fei WANG, Zhi LIU, Han GAO, Huiji WANG. Correction of vignetting images based on Retinex-Net network model[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(7): 929

    Download Citation

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

    Category: Research Articles

    Received: May. 25, 2023

    Accepted: --

    Published Online: Jul. 23, 2024

    The Author Email: Zhi LIU (liuzhi@cust.edu.cn)

    DOI:10.37188/CJLCD.2023-0194

    Topics