結合多重特徵抽取法及支持向量機於軟性面板之瑕疵檢測：以雙穩態膽固醇液晶顯示器為例

Defect Inspection for Flexible Displays Using Multiple Feature Extraction Methods and Support Vector Machine: A Case Study on Bi-Stable Cholesteric Liquid Crystal Display

雙穩態膽固醇液晶顯示器 （bi-stable cholesteric liquid crystal display,BS-ChLCD） 是一種具有輕、薄、可撓曲、省電、具記憶性等特點的新型軟性顯示器。由於BS-ChLCD目前正處於量產研發階段，以致於面板在Roll-to-Roll製程中，將無法避免人為或機台等因素而產生瑕疵，進而影響到面板的顯示效果。因此，本論文利用機器視覺及機器學習理論針對BS-ChLCD 進行表面瑕疵檢測系統開發。此系統主要由三個程序所構成：影像前處理、訓練程序及檢測程序。首先，本系統將一張原始影像切割成數張子影像，接著進行瑕疵分類器的訓練。但待測影像並非呈現完全水平的狀態，而這樣的誤差將導致後續重建背景紋路時會發生問題。因此，在進入訓練及檢測程序之前，必須先將影像調校為水平，以利於後續檢測流程。在訓練程序中，利用紋理 （texture）、變異數 （variance）、奇異值分解 （singular value decomposition, SVD） 及主成分分析（principal component analysis, PCA） 等特徵抽取方法，對每張分割後的子影像經過特徵抽取後，便形成一組特徵向量。最後利用支持向量機 （support vectormachine, SVM） 來建構出一個瑕疵檢測模型。檢測程序中，將一張原始影像經過前處理後，以相同的特徵抽取方式輸入至SVM 瑕疵檢測模型，便可判斷該張子影像是否具有瑕疵，若存在瑕疵，則將這張子影像加以標記；若無瑕疵，子影像進入背景紋路的斷線偵測系統，直到所有子影像都經過一系列的瑕疵判別之後，即完成一張影像的檢測。本論文在實驗中所採用之影像，皆為工研院影像顯示中心所提供之實際BS-ChLCD 影像。實驗結果顯示，本論文所提出之BS-ChLCD 瑕疵檢測系統，瑕疵檢測率高達99.04%。

Bi-stable Cholesteric Liquid Crystal Display (BS-ChLCD) is a newkind of flexible display. It has good properties of low power consumption,flexibility, and slimness. Due to different physical factors, such asmachine breakdown and particles, various defects occur on the surfaces ofBS-ChLCD panels during the roll-to-roll manufacturing process. Wepropose, in this paper, a system which can automatically detect the surfacedefects from the images acquired from real BS-ChLCD panels. Thesystem is composed of three layers, image preprocessing, feature extraction,and defect inspection layers. The image preprocessing layer isresponsible for the segmentation of an input image. After segmenting theinput image into non-overlapping subimages, the subimages are then sentinto the second layer for feature extraction, one subimage at a time. Weemploy three feature extraction methods in this study, including singularvalue decomposition (SVD), variance, and principal component analysis(PCA). After extracting the features from a subimage, the subimage canbe represented by a feature vector, which will be fed into the third layer forfurther classification. In the defect inspection layer, a novel machinelearningmethod called support vector machine (SVM) is used as the defectdetector. The input to the SVM classifier is the feature vectors, and theoutput is either normal or defective. Once a feature vector is classified asdefective, the corresponding subimage is defective. Moreover, the defectdetection task on the input image is accomplished once the class labels ofthe subimages are obtained. According to our experimental results, theproposed system is able to achieve a high defect detection rate of over99%.