- 學位論文
影像處理及類神經網路於晶圓缺陷分析之應用
Automatic Wafer Defects Analysis Using Image Processing Techniques and Neural Networks
摘要
半導體晶圓缺陷的分析研究運用影像判讀的需求性日益增加，而目前仍有製程尚需仰賴人工目視判讀分群，不只曠日費時，此外這與工程師分群判讀技術與專業訓練成熟度也密切相關，因此極需客觀的資訊工具數位化整個過程，以最迅速的方式來給予工程人員建議資訊。 本研究的目的主要運用影像技術及類神經網路自動化分析晶圓缺陷，首先將原影像轉成灰階圖再運用Otsu 演算法自動尋找最佳閥值（Optimal Threshold）進行影像分割，並結合形態學（Morphology）的技巧來消除雜訊，保留特徵型態，接著根據取得的輪廓資訊去計算影像形狀特徵值及將此輪廓套回原影像取得內部結構的特徵值，接著利用徑向基底函數類神經網路（Radial Basis Function Network , RBFN）快速學習的特性對晶圓缺陷特徵值做辨識，更有效率的辨識分群大量晶圓影像資料，目前研究成果所提出的晶圓缺陷影像辨識架構，辨識率約在83%~87%之間。
並列摘要
Analysis of semiconductor wafer is increasingly in need of applying imaging determination, yet present process still relies partly to human visual determination for categorization. It takes time and the results are closely related to the categorization and determination techniques and professional training maturity of engineers, so it is necessary to have a digitized the process with objective information instrument to provide information and suggestion to engineers in the fastest way. In this research, we applied imaging techniques and artificial neural network to analyze wafer defects automatically. In the first, the raw image is converted into gray-level image, then use Otsu computation to search for optimal threshold automatically to do image division. They Morphology techniques to eliminate noise, but retain the characteristic pattern. The obtained profile information is based to compute the image shape feature and then overlap the profile back to the original raw image to obtain the feature of internal structure. Thereafter, the fast learning features of Radial Basis Function Network (RBFN) is used to identify the wafer defects features. It enables efficiently differentiate categorized massive wafer image information. The identification rate of the Wafer Defects Image Identification Structure as presented through this research is at 80% to 85%.