在TFT-LCD產業，對於面板點燈檢查上，有使用機器視覺及Rule base演算法開發的面板自動瑕疵檢查系統，把產品分為正常品及瑕疵品。瑕疵品會讓人員進行目視檢查，降低系統判定錯誤的數量。由於面板表面有多種假性瑕疵現象，以及面板尺寸大型化，單一片面板需檢查的面積增加等因素，會增加系統判定錯誤的機率，進而增加人員覆判的成本。
本研究中，使用多個卷積神經網路的模型檢測方法。利用Rule base演算法所判定為瑕疵的局部特徵圖檔，經由人員標示為陰性或陽性，讓模型進行學習及驗證後，進行自動覆判，改善面板檢測的準確率以及作業效率。本文選用三種模型進行比較，驗證模型的適用性。針對面板圖檔特有的訊號特性，進行影像前處理，研究有無處理的正確率差異性。為減低資料集數量不夠的問題，採用了資料擴增的方法，增加不同方向特徵數量。為減少訓練集、驗證集在資料排列上所造成的訓練偏差，使用了k-Fold交叉驗證方法，研究不同排列方式的模型穩定性。透過本研究的模型訓練及驗證方法，可用於自動化面板瑕疵檢查覆判作業。

Automatic defect inspection systems based on machine vision and rule-based algorithms are used for panel lighting inspection in the TFT-LCD industry. These systems classify products into normal and defective products. Defective products are visually inspected further by humans to reduce the number of errors identified by the system. Due to a variety of false defects on panel surfaces and their large sizes, including an increase in the area to be examined for a single panel, the cost associated with both the probability of system errors and personnel re-judgment increases.
This study developed a model detection method based on multiple convolutional neural networks. The local characteristic image files, which were judged as defective by the rule-based algorithms, were marked as either negative or positive by the inspection personnel. Subsequently, the model was verified by learning to conduct automatic re-judgment and improve the accuracy and operational efficiency of panel detection. Three models were compared to verify the applicability of the proposed model. The special signal characteristic image files of the panel were pre-processed to study differences in the accuracy rate between the processed and non-processed signals. To address the problem of insufficient data sets, a data amplification method was adopted to increase the number of characteristics in different directions. Further, to reduce the training deviation caused by the data arrangement of the training set and the validation set, the K-fold cross-validation method was used to study the model stabilities of different arrangements. The model training and validation method presented in this study can be applied for automatic panel defect inspection and re-judgment.

Batchelor, B., Hill, D and Hodgson, H. (1985). Automated visual inspection.

D. M. Tsai, and W. C. Li., (2011). Defect inspection in low-contrast LCD Images using Hough transform-based Non-stationary line detection., IEEE Transactions on Industrial Information, Vol.7, No.1, pp. 136-147.

Gerald C. Holst. (1996). CCD Arrays, Cameras, and Displays. SPIE Optical Engineering Press and JCD Publishing, Washington DC.

Girshick, R., Donahue, J., Darrell, T., Malik, J. UC Berkeley, and ICSI., (2014). A rich feature hierarchies for accurate object detection and semantic segmentation In CVPR.

Girshick, R. (2015). Fast R-CNN, IEEE International Conference on Computer Vision

Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun., (2015). Deep Residual Learning for Image Recognition., In arXiv:1512.03385v1 [cs.CV] 10 Dec.

Karen, S., Zisserman, A. (2015). Very Deep Convolutional Networks for Large-Scale Image Recognition, University of Oxford.

Krizhevsky, A., Sutskever, I. and Hinton, G.E. (2012). ImageNet classification with deep convolutional neural networks. In NIPS.

LeCun, Y., Haffner, P., Bottou, L., & Bengio, Y. (1999). Object recognition with gradient-based learning. In D. Forsyth, J. Mundy, V. di Gesu, & R. Cipolla (Eds.),shape, contour and grouping in computer vision(pp.319-345). Springer Verlag.

Li-Fei Chen, Chao-Ton Su, and Meng-Heng Chen., (2009). A Neural-Network Approach for Defect Recognition in TFT-LCD Photolithography Process., IEEE TRANSACTIONS ON ELECTRONICS PACKAGING MANUFACTURING, VOL. 32, NO. 1, JANUARY.

Mingxing Tan., Quoc V. Le. (2019). Rethinking Model Scaling for Convolutional Neural Networks＂arXiv:1905.11946v1 [cs.LG] 28 May.

N. K. Park, and S. I. Yoo., (2009). Evaluation of TFT-LCD defects based on human visual perception, Displays, Vol. 7, No.1, pp.1-16.

S. K. S. Fan, and Y. C. Chuang., (2010). “Automatic detection of Mura defect in TFT-LCD based on regression diagnostics,＂Pattern Recognition Letters, Vol. 33, No. 15, pp. 2397-2404.

Vernon, D. (1991). Machine vision automated/visual inspection and robot vision, New York,1991.

周尚達, (2007). 應用小波轉換於TFT-LCD之Mura瑕疵檢測,碩士學位論文,成功大學。

陳思宏,(2006). 彩色濾光片 Mura 瑕疵之瑕疵檢測，碩士學位論文，義守大學。

陳茂旗,(2011).電腦視覺應用於TFT-LCD基板之表面瑕疵檢查,碩士學位論文,長庚大學。

張瑞顯,(2005). 應用線性回歸診斷法於液晶顯示器Mura缺陷自動化檢測之設計與實現,碩士學位論文，義守大學。

黃思博,(2013). 應用自動化光學檢測技術於 TFT-LCD模組瑕疵辨識系統之開發與研究，碩士學位論文，台灣科技大學。

曾彥馨,(2004). 應用機器視覺於 TFT 面板之表面瑕疵檢與分類.,碩士學位論文,元智大學。

蔡佳航,(2011). 應用小波轉換與模糊支向機於TFT-LCD Cell製程自動化瑕疵分類之研究，碩士學位論文，明新科技大學。