建構以約略集合理論為基礎之TFT-LCD自動光學檢測瑕疵分類模型

Abstract

自動光學檢測是目前發展迅速且逐漸被製造商採用的一種品質檢驗技術，透過光學儀器取像後搭配影像處理演算法對檢測的對象進行檢驗，取代人力檢測之不足。目前在關於液晶面板(TFT-LCD)自動光學檢測的文獻中，僅有極少數的文章針對面板上的壞點瑕疵進行研究，然目前在業界皆以壞點數作為面板品質的分級標準。有鑑於此，本研究以面板上的壞點瑕疵為檢測對象，透過本研究所特別設計的影像前處理方法，應用約略集合理論(Rough Set Theory，RST)建構一套TFT-LCD面板自動光學檢測的壞點瑕疵分類模型。本研究利用台灣某電腦品牌商所提供之面板真實影像作為驗證資料，經過實驗後發現，本研究所設計的TFT-LCD瑕疵分類模型，對於這些驗證資料平均有99.5%的整體分類準確率，表示本研究所提出之方法的有效性。本研究亦與過去文獻中所使用的分類模型進行效能比較，包含支持向量機分類模型與提升決策樹分類模型，研究結果顯示本研究所設計的RST分類模型在壞點瑕疵的分辨能力與檢測時間方面的表現皆優於這兩種方法所建立的分類模型。

Automatic Optical Inspection (AOI) is a rapid-developing quality inspection technology that has been gradually adopted in practice. Through images taken by optical instruments along with image-processing algorithms, one can inspect the objects of interest automatically to replace manual inspections. Among the literature regarding the application of AOI in TFT-LCD panels, very few studied the defective pixels. However, the number of defective pixels is usually used as a standard to grade TFT-LCD panels. Therefore, this study aims at constructing a classification model based on the existence of defective pixels on TFT-LCD panels. Special image pre-processing processes were designed to emphasize different types of defective pixels. A classifier based on rough set theory (RST) was designed to classify the images with defective pixels from the ones without. Real images of TFT-LCD panels provided by a computer manufacturer in Taiwan were used to test the effectiveness and efficiency of the proposed approach. The results of numerical experiment show that the proposed approach has average of 99.5% classification accuracy. The RST classifier designed by the proposed approach was also compared with the ones based on support vector machine and boosted decision trees, respectively, used by previous studies. The results show that the RST classifier outperformed the other two classifiers.