基於相似程度之自適性共振1號模型之自動缺陷偵測與分類

Abstract

本論文將建立一套自動化缺陷偵測與分類系統，在缺陷偵測的部分，利用MAD的方法將待測影像與參考影像對齊，其中參考影像為不包含缺陷的影像。接著，我們將對齊後的待測影像與對齊後的參考影像相減，並進行二值化的程序以獲得二值化差異影像。最後考量到在相減影像中會有一些微小的零星雜訊，我們設定一個最小相連接像素點個數閥值來過濾這些面積較小的零星雜訊，最後，從待測影像當中擷取出缺陷影像以進行缺陷分類。 缺陷分類的部分，基於自適性共振理論1號模型的架構，其優勢在於能夠同時滿足模型的穩定性與可塑性。本論文發現其分類模型有輸入缺陷影像順序不同而分類結果也不同之缺失，為了改善此缺失，本論文建立一套基於相似程度之自適性共振理論分類演算法進行缺陷分類的程序，此演算法保留自適性共振理論1號模型之可塑性的優點，並具有相當的分類能力且改善了分類結果會因為輸入順序不同而改變的缺點。

In general, a reliable and automatic semiconductor fabrication processes is of great importance to product yield and cost reduction. In the past, we made use of human vision to do die defect detection and classification, which is hindered by the easy fatigue and fuzziness of human eyes and the decision difference between inspectors. In this thesis, we develop a vision-based automatic defect classification system. In our defect detection component, we apply the MAD method to align the test image with the reference image. To acquire the binary defect images, we subtract the test image from the reference image, then we convert the difference image into the binary image by setting a threshold. Moreover, we removed the scattering noises by setting a minimum number of connected noisy pixels required. Finally, we extract all defects in the test image in order to perform the defect classification. For defect classification, we revise the ART 1 model, which still can retain the stability and the plasticity dilemma. We have found that ART 1 exists an intolerable shortcoming: output is dependent on the ordering of input sequence applied. To remedy this disadvantage, we derive the similarity based ART 1 model which can obtain high classification accuracy and independent on the ordering of input patterns.