應用於晶圓圖分群與分類具旋轉不變性之相似度量

Abstract

半導體製程中有上百個步驟，當其中任何機台出現問題，很有可能對良率造成影響。過去依賴有經驗的工程師以人工辨識的方法，藉由晶圓瑕疵圖辨別製程中發生問題的原因。如此不僅效率較低，且個人判別的標準並不一致，無法給予一個客觀的判定。本論文提出一客觀的晶圓圖相似度指標，以利晶圓圖分群及分類。晶圓尺寸、旋轉、位移及圖形大小是在定義晶圓圖間相似度時會面對到的問題。我們以正規化解決了晶圓尺寸不一的問題，並以具有旋轉不變性的Zernike動差作為特徵解決了影像旋轉的問題，最後再以Zernike動差特徵空間之歐幾里得距離作為不相似度指標。實驗證明Zernike動差特徵於群聚樹分群、加權式KNN分類，及相似比對皆有顯著效果。

There are hundreds of steps in a semiconductor manufacturing process. If any problems occur in these steps, it might reduce the yield of the wafers. To determine the causes of the yield loss, visual recognition of wafer bin map patterns by experienced engineers is a common practice in present wafer manufacturing industry. It’s neither efficient nor an objective method with inconsistent-recognitions holding by each engineer. In this thesis, we propose an objective similarity measure for wafer bin map classification and clustering. There are four issues in devising wafer map similarity measure including wafer size, image rotation, image translation, and the size of the pattern. We solved the wafer size issue with normalization, and overcame the rotation issue with a feature extraction method called Zernike moment, which ensures the rotation-invariant property. After the feature extraction, we use the Euclidean distance in Zernike moment space as a dissimilarity measure. The experiments showed that the Zernike moment feature had good performances in linkage-tree clustering, distance-weighted KNN classification, and query matching.