應用遞增式支持向量機於TFT-LCD陣列電路工程閘電極光罩線中瑕疵辨識

Applying Incremental Learning-based SVM to Gate-Electrode Mask Defect Classification for the Inline Inspection of TFT-LCD Array Engineering

林佑駿(You-Jun Lin)

指導教授 ： 劉益宏

中原大學/工學院/機械工程研究所/碩士(2008年)

https://doi.org/10.6840/CYCU.2008.00368

摘要

對於目前的面板廠商而言，最重要的目標之ㄧ是提高面板的良率，以降低生產成本並增加公司獲利。為了增加良率，多數的公司都設有檢查部門，主要是當製程進行到某一個階段時進行人工的瑕疵分類工作。為了降低人為因素的影響以及增加瑕疵分類的速度，整個檢查系統自動化是必須的。然而，瑕疵的變異非常大，基於某一個時期的瑕疵資料所訓練出來的分類器，對於未訓練過的瑕疵資料並不能提供一個很好的分類能力。因此，建立一個線上重新訓練的機制是必須的。
本論文提出『基於遞增式學習之支持向量機』的方法來解決這個問題。其可以利用既有的支持向量與分類錯誤的新進資料作結合產生新的訓練資料，再應用該資料去重新訓練一個最佳分離超平面。該方法不但可以以減少訓練資料量，也可以減少重新訓練所花費的時間。更重要的是，它不但能保留原有資料的資訊，且能適應於新進的資料，使分類器具有學習的能力，進而改善分類率。本研究針對陣列工程中、閘電極光罩中光微影製程瑕疵影像進行辨識，包含常見的瑕疵，例如『刮傷』、『閘電極-儲存電容短路』、『塗佈異常』、『異物』等。實驗的瑕疵照片皆由面板廠所提供。實驗結果顯示，本論文所提出的『基於遞增式學習之支持向量機』可以達到 95% 以上的辨識率。

關鍵字：薄膜電晶體液晶顯示器、遞增式學習、支持向量機、瑕疵分類

關鍵字

遞增式學習 ； 瑕疵分類 ； 薄膜電晶體液晶顯示器 ； 支持向量機

並列摘要

For current TFT-LCD manufacturer, one of the most important goals is to enhance the yield rate and reduce the production cost. To enhance the yield rate, most companies have set up the inspection departments to manually perform the task of defect classification. A fully automatic inspection system is required in order to reduce the effects due to human factors and speed up the defect classification. However, the variation of defects is very large. That is, the classifier trained with the defect data collected with a specific period can not provide a good classification accuracy for those do not included in the training phase. Therefore, it is necessary to develop an online re-training model.
As a classifier, support vector machine (SVM) has shown to be superior to the multi-layer neural networks trained by principle of empirical risk minimization. The learning strategy of SVM is based on the principle of structural risk minimization. Therefore, SVM can provide good generalization performance for unseen defect data. However, as mentioned above, the variation of defects is large. It can be expected that SVM will still suffer from the problem that some unseen defects would be misclassified. Therefore, how to develop an online training algorithm for SVM would be the key issue for achieving high TFT-array-process defect classification rate. This thesis focuses on this.
In this thesis, an incremental learning based support vector machine (IL-SVM) is proposed to deal with this problem. IL-SVM is able to combine the existing support vectors and the newly coming data that are misclassified to re-train an optimal separating hyperplane (OSH). This method can not only reduce the number of training data required, but also reduce the time for re-training the SVM. What is the most important, IL-SVM can not only preserve the information of original training data, but also adapt to the newly coming data, thus improving the classification rate. For achieving the best cross-validation rate, the radial basis function (RBF) is adopted as the kernel function for SVM.
This study aims at recognizing the types of the defect images captured in the photolithography process of gate-electrode engineering in TFT array process. The defects include “scratch”, “connection of gate electrode and capacity storage”, “abnormal resist coating”, and “foreign object”. They are commonly-seen defects in TFT array process, and are critical to TFT panels. The goals of real-time equipment diagnosis and maintenance can be achieved if the types of the defects can be timely recognized because each kind of defect has each cause of generation. As a result, the yield rate can be improved.
In additional to the classifier design, this paper also discusses the feature selection problem. To obtain better representation of a defective image, the principal component analysis (PCA) algorithm is used to transform the input vector into a new one which has a lower dimension. The experiments are conducted on a set of real defect pictures provided by a TFT-LCD manufacturer. Experimental results show that the proposed IL-SVM is able to achieve a high defect recognition rate of over 95%. The results also indicate that PCA can further improve the classification accuracy.

Keywords: thin film transistor liquid crystal display (TFT-LCD), incremental learning, support vector machine (SVM), defect classification, principal component analysis.

並列關鍵字

support vector machine (SVM) ； incremental learning ； principal component analysis ； thin film transistor liquid crystal display (TFT-LCD) ； defect classification

參考文獻

[26] 林思賢，”應用視覺及資料探勘技術於TFT-LCD陣列電路工程線中瑕疵辨識之研究-源/汲電極光罩瑕疵自動分類系統開發”,私立中原大學機械工程研究所，2006年

[3] R. Jain, R. Kasturi and B. G. Schunck, Machine Vision, McGraw-Hill, New York, 1995.

[4] Woo-Seob Kim, Dong-Min Kwak, Young-Chul Song,Doo-Hyun Choi and Kil-Houm Park, “Detection of Spot-Type Defects on Liquid Crystal Display Modules,” Key Engineering Materials, vols. 270-273, pp. 808-813, 2004.

[6] Chi-Jie Lu and Du-Ming Tsai, “Automatic Defect Inspection for LCDs Using Singular Value Decomposition,” International Journal of Advanced Manufacturing Technology, vol. 25, pp. 53-61, 2005.

[7] Yumi Mori, Kosei Tanahashi and Satoshi Tsuji, “Quantitative evaluation of mura in liquid crystal displays,” Optical Engineering, vol. 43, no. 11, pp. 2696-2700, November 2004.