應用奇異值分解法(SVD)與獨立成份分析法(ICA)於薄膜電晶體液晶顯示面板之表面瑕疵檢測

Automatic defect inspection for patterned TFT-LCD panel surfaces using singular values decomposition and independent component analysis

呂奇傑(Chi-Jie Lu)

指導教授 ： 蔡篤銘 博士

元智大學/工程學院/工業工程與管理學系/博士(2005年)

摘要

薄膜電晶體液晶顯示器（Thin Film Transistor Liquid Crystal Displays, TFT-LCDs）具備輕、薄及省電之特性，近年來廣為應用於消費性電子產品上。目前TFT-LCD面板的瑕疵檢測仍然仰賴人工進行檢測，人工檢測不僅需要昂貴的人工成本，並且無法進行線上即時檢測，本研究導入機器視覺於TFT-LCD面板之表面瑕疵檢測，以解決人工檢測所面臨的限制以及降低其所帶來的成本。本研究基於影像重建（image reconstruction）之架構提出兩種瑕疵檢測方法來進行TFT-LCD面板之表面瑕疵檢測，檢測之對象為扭轉向列型（Twisted Nematic, TN）與多象限垂直配向型（Multi-Domain Vertical, MVA）等二類TFT-LCD面板，檢測之瑕疵設定為人工不易檢測之微觀瑕疵，包括孔洞(pinhole)、刮痕(scratch)以及粉塵(particle)。
本研究所提之第一種方法為利用奇異值分解法（Singular Value Decomposition, SVD）對TN型TFT-LCD面板上的微觀瑕疵進行檢測。TN型TFT-LCD面板的表面在影像中具單純之結構式紋路，正交的垂直信號線（data line）與水平閘線（gate line）規律的構成面板影像的背景紋路。我們先使用奇異值分解法於輸入的TFT-LCD影像以獲得一個對角矩陣。對角矩陣中的值稱為奇異值（singular values)，再選擇較大的奇異值來代表TFT-LCD影像的背景紋路後，我們可以去除此代表背景紋路之奇異值然後重建影像。在重建後的影像中，TFT-LCD影像中的規律性背景紋路將被去除而只保留瑕疵。
本研究所提之第二種方法為利用獨立成份分析法（Independent Component Analysis, ICA）對TN型以及MVA型TFT-LCD面板進行瑕疵檢測。在MVA型TFT-LCD面板之影像中，垂直信號線、水平閘線、共通線（common line）與具複雜結構的薄膜電晶體（TFT）等元件規律的構成面板影像之背景紋路。由於元件中之薄膜電晶體具複雜結構，使得奇異值分解法無法良好的應用於MVA型TFT-LCD面板之瑕疵檢測上。在方法二中，我們先利用獨立成份分析法將TFT-LCD影像分離成獨立成份（Independent Components, ICs）與反混合矩陣（de-mixing matrix）的組合，再選擇適當的獨立成份來代表TFT-LCD影像中的規律性紋路結構後，原始的反混合矩陣將被修正以產生新的反混合矩陣，新的反混合矩陣將不包含TFT-LCD面板中規律紋路之資訊。最後，利用此一新的反混合矩陣對待測的TFT-LCD面板進行影像重建，重建後的影像中將只保留瑕疵並濾除TFT-LCD面板中之規律性紋路。
本研究所提的方法一僅適用於偵測TN型TFT-LCD面板之微觀瑕疵，而無法應用於檢測MVA型TFT-LCD面板。方法二則可以同時用於偵測TN型以及MVA型TFT-LCD面板上之微觀瑕疵。然而，方法二僅能偵測出面板上有無瑕疵存在，而無法明確的顯示瑕疵之形狀。反之，方法一則不僅能夠偵測出瑕疵的位置，同時也能保留瑕疵的形狀與大小。
本研究所提的兩種影像重建為基之檢測方法均無需參考影像之比對，也不需進行紋路特徵之擷取，因此可以避免樣板比對法（template matching）以及特徵擷取法（feature extraction）的所面臨限制與缺點。實驗結果顯示，本研究所提的第一種以奇異值分解法為基之方法可以有效的檢測出TN型TFT-LCD面板上之微觀瑕疵以及瑕疵位置與形狀。第二種以獨立成份分析法為基之方法則對於TN型以及MVA型TFT-LCD面板上之微觀瑕疵均具有良好偵測效果。

關鍵字

薄膜電晶體液晶顯示面板 ； 表面瑕疵檢測 ； 奇異值分解法 ； 獨立成份分析法 ； 影像重建

並列摘要

Thin Film Transistor Liquid Crystal Displays (TFT-LCDs) have become increasingly popular as display devices. For ensuring the display quality of TFT-LCD panels, human visual inspection is the most commonly used method for inspecting the surface defects in LCD panels. In this research, two machine vision approaches based on image reconstruction are proposed for the inspection of surface defects in patterned TFT-LCD panels. Micro defects including pinhole, scratches and particles in Twisted Nematic (TN)-type and subtle defects such as stain and particles in Multi-Domain Vertical (MVA)-type LCD panel surfaces are the main focus of this research.
In the first approach, a singular value decomposition (SVD) based image reconstruction scheme is proposed to detect micro defects in TN LCD panel surfaces. The geometrical structure of a TN LCD panel surface involves only repetitive simple horizontal gate lines and vertical data lines. The TN LCD image is a textured image of simple structural pattern. Taking the pixel image as a matrix, the singular values on the decomposed diagonal matrix represent different degrees of detail in the textured image. By selecting the dominant singular values that represent the background texture of the patterned surface and reconstructing the matrix without the selected dominant singular values, we can eliminate periodical, repetitive patterns of the textured image, and preserve the anomalies in the reconstructed image.
In the second approach, an independent component analysis (ICA) based image reconstruction scheme is presented to detect micro defects in both TN and MVA panel surfaces. ICA is applied to a faultless training image to determine the de-mixing matrix and its corresponding independent components (ICs). After identifying a set of proper ICs that represent the global structure of the LCD panel, the corresponding rows of the de-mixing matrix for the identified ICs are then replaced with a de-mixing row associated with the least global background information to reform the de-mixing matrix. The reformed de-mixing matrix containing only uniform information of the panel image is used to reconstruct the TFT-LCD image under inspection.
In the comparison of the two proposed TFT-LCD inspection schemes, the first SVD-based method can only be used for inspecting the TN LCD panels that contain a simple structural pattern. The second ICA-based method can be used for inspecting both TN and MVA LCD panels that contain complicated texture structures. However, the ICA-based method cannot preserve the size and shape of a detected defect in the reconstructed image. It can only detect the presence of a defect in the inspection image. The SVD-based approach not only can detect the location of a defect in the inspection image, but also well preserve the size and shape of the detected defect in the reconstructed image.
The two proposed methods do not rely on the design of textural features to detect local defects, nor does it require a template image for comparison. They can ease the limitations of feature extraction and template matching methods. The experiments reveal that the first proposed SVD-based method has shown promising results for micro defect inspection in TN LCD panels. The second proposed ICA-based method is effective for detecting the subtle defects in both TN and MVA LCD panels.

並列關鍵字

TFT-LCD panels ； Defect inspection ； Singular value decomposition ； Independent component analysis ； Image reconstruction