Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 11, 1463(2024)

Development of AOI inspection of Mura defects on TFT-LCD surface

Zekang CHEN1, Yi SHEN2, Chenyang ZHAI1, Chenyao DONG1, and Shuangxi WANG1、*
Author Affiliations
  • 1College of Engineering,Shantou University,Shantou 515063,China
  • 2Guangdong Provincial Key Laboratory of Automotive Display and Touch Technologies,Shantou 515041,China
  • show less
    Figures & Tables(15)
    Development process of liquid crystal display technology[6]
    Manufacturing process of TFT-LCD[6]
    Schematic diagram of common Mura defects[16]
    Comparison of three-dimensional maps of gray values of Mura defects with different filtering effects[38]
    Comparison of brightness correction results[29]
    Common Mura defect manifestations on LCD screens
    Screen surface Mura defect recognition results sheets[40]
    Field of view simulation for defect detection[45]
    TFT-LCD Mura defect visual inspection method in multiple backgrounds[52]
    Localizing texture defects with the proposed MSCDAE model on marble samples(first row),LCD panel samples(second row),and fabric samples(third row).(a)Original defect images;(b)Residual map of the first pyramid layers;(c)Residual map of the second pyramid layers;(d)Residual map of the third pyramid layers;(e)Final comprehensive results[56].
    Computational cost of the method based on the OS-ELM and that based on the ELM[60]
    • Table 1. Comparison of Mura defect image feature extraction methods

      View table
      View in Article

      Table 1. Comparison of Mura defect image feature extraction methods

      类别具体方法优点局限参考文献
      几何特征基于形态学操作的移动屏幕检测系统不依赖于颜色和亮度信息,计算简单且具有较强的抗噪声能力对图像表面噪声较为敏感、计算量较大44
      颜色、灰度特征比较屏幕图像在不同标准下的色彩表现较强的鲁棒性,对图像的旋转、平移、尺度变化等不敏感灰度空间下图像处理效果较好,其他标准下的图像伴随较多的干扰46
      统计特征基于人眼感知的特征量化计算效率高、可解释性强需要手动选择参数,对复杂缺陷的 鲁棒性较差47
    • Table 2. F1-measures of different methods on five types of textures59

      View table
      View in Article

      Table 2. F1-measures of different methods on five types of textures59

      数据集织物瓷砖水泥壁纸液晶屏
      PHOT0.0010.2620.0110.1320.452
      TEXEMS0.0010.4110.2850.1420.556
      AE0.3520.1940.2170.3910.678
      OCGAN0.0010.2420.1670.0850.603
      MS-FCAE0.7270.3400.0790.1120.739
      AFEAN0.8410.6240.7340.8520.867
    • Table 3. Average accuracy of each model55

      View table
      View in Article

      Table 3. Average accuracy of each model55

      模型类型无噪声缺陷图像带噪声缺陷图像缺陷图像去噪
      VGG1682.54±2.4776.3±5.2381.72±4.83
      ResNet5080.10±3.9371.02±3.9479.43±3.85
      InceptionV384.22±3.7377.85±5.3283.38±5.17
      Transfer Ensemble91.06±2.3786.11±2.4890.13±1.89
    • Table 4. Comparison of Mura defect detection methods based on deep learning

      View table
      View in Article

      Table 4. Comparison of Mura defect detection methods based on deep learning

      分类类别模型举例优点局限
      有监督学习目标检测YOLOYOLO模型检测速度快,样本标注简单需要大量的缺陷样本并进 行人工标注,必须解决正负样本不均衡、不同尺度大小的缺陷样本识别效果不均 等问题
      目标分类基于卷积神经网络CNN分类、DCANet等只对整幅图像中的目标区域进行标注,训练和测试花费时间少
      目标分割Mask R-CNN等训练与测试花费时间少,可精准分割目标
      无监督学习基于生成对抗网络(GAN)的缺陷检测GAN、Res-unet-GAN等不需要过多的缺陷样本即可进行训练,无需标注对缺陷重构区域精度要求较高,容易出现过拟合
      基于AE的缺陷检测MS-FCAE、AFEAN
      迁移学习

      基于多任务学习模型

      基于对抗性领域自适应模型

      共享编码器/解码器架构

      DANN

      有效利用源任务中的先验知识,提高目标任务的学习效率和性能源任务与目标任务之间的相似性要求较高,否则容易出现负迁移
    Tools

    Get Citation

    Copy Citation Text

    Zekang CHEN, Yi SHEN, Chenyang ZHAI, Chenyao DONG, Shuangxi WANG. Development of AOI inspection of Mura defects on TFT-LCD surface[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(11): 1463

    Download Citation

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

    Category: Liquid Crystal Optics

    Received: Aug. 14, 2024

    Accepted: --

    Published Online: Jan. 3, 2025

    The Author Email: Shuangxi WANG (sxwang@stu.edu.cn)

    DOI:10.37188/CJLCD.2024-0235

    Topics