以自動化視覺檢測系統為基礎之瑕疵分類研究

A Study of Defects Classification Based on Automatic Visual Inspection System

王建智(Chien-Chih Wang)

指導教授 ： 江行全

元智大學/工程學院/工業工程與管理學系/博士(2001年)

摘要

本論文之研究為在以自動化視覺系統的架構下，發展新的分類模型並應用在電子產品的瑕疵分類。探討影響整個分類系統效率的關鍵因素，主要有特徵變數的選取以及分類演算法的設計，針對這兩部份，本研究發展出新的方法與程序，並以實際收集的銲錫點資料和文獻上之資料為例，驗證本研究所提方法。
就特徵變數的選取部份來說，本研究分別針對單變數特徵變數與多變數特徵變數子集的選取，發展出新的演算法。
(I) 單特徵變數選取：提出以FSI(Feature Selection Index)指標來量化兩群體間重疊區域的大小。在常態分配下推導出兩群體重疊區域面積與FSI指標的關係，藉此作為最佳單特徵變數選取的依據。並推廣FSI而以FSI之SN比的望大公式作為多群體之最佳單特徵變數選取之指標。
(II) 多特徵變數選取：整合實驗設計與多變量變異數分析的技巧，來進行最佳多特徵變數選取。首先以單變量特徵變數法淘汰區別力較差的特徵變數，然後將剩下的特徵變數設計一解析度為III的PB設計，經由此過程可以大量減少特徵變數搜尋時間。接下來，進行每一實驗組合在不同群體下的MANOVA，計算Pillai統計量來作為反應值。最後則透過統計分析，得到一組最佳特徵變數。
就分類演算法的設計，本研究分別對於有母數與無母數分類法提出相對應的新策略。
(I) 母數演算法：提出最適化貝氏分類程序，解決非高斯分配下採用貝氏分類所產生的高分類誤差率的缺點。經實際資料驗證，結果確實優於傳統貝氏分類。
(II) 無母數演算法：配合修正之樹狀分類結構，提出以群體重疊區域資料為基礎下的最適分類規則。以Normal分配、Gamma分配和Weibull分配下分別就右偏、對稱與左偏的組合討論。經模擬驗證，本研究所提方法優於貝氏分類法。
將論文所提方法，整合成兩分類系統。I：整合以重複區域為基礎之特徵變數選取和順序分類樹，II：整合多變量特徵變數選取和最適貝氏分類程序。並將應用在銲錫點的分類問題上，經分析以系統I來說，分類正確率之信賴區間為(98.5%, 100%)，優於貝氏分類法與區別函數分類法。以系統II來說，平均分類正確率為94%，優於傳統貝氏分類法、區別函數分類法和最鄰近分類法。而就文獻上的OCR資料來說，本論文所提方法之正確分類率亦高於文獻結果。

關鍵字

自動化檢測 ； FSI ； 實驗設計 ； MANOVA ； 瑕疵分類 ； 銲錫 ； OCR

並列摘要

This research is based an automatic visual system to develop a new classification model and to apply it to the printed circuit board (PCB) defects classification. The feature selection and classifier design are key factors for the classification. To address these two aspects, this research proposed a new algorithm and procedure, and using the solder joints for a PCB from a monitor manufacturer in Taiwan and a set of OCR data to verify these methods.
Two new algorithms for feature selection were studied：
(I) Single Feature Selection：The feature selection index (FSI) was designed to measure the overlapped region for two different groups of data. Under the normal distribution assumption, it was proved that the relationship between two different overlapped region and the FSI values, and it could used to select the optimal single feature variable. Furthermore, the-larger-the-better SN ratio was used to calculate the multiple groups’ FSI to get the optimal feature variable.
(II) Multiple Features Selection：To integrate DOE and MANOVA techniques to select the optimal multiple feature variables. First, the single feature variable selection algorithm was adopted to eliminate poor discrimination feature variables. Then the Plackett-Burman (PB) resolution III design was constructed for the selection of remaining feature variables. Using MANOVA technique, calculate Pillai statistic as the response for the PB design of experiment. Finally, execute statistical analysis to obtain the optimal multiple feature variables for multiple groups.
Two new algorithms of classification were studied：
(I) Parametric Classification Algorithm：An adaptive Bayesian classification procedure was adopted to solve high classification error rate problem when utilizing the normal Bayesian classification method under non-normal distributed data.
(II) Nonparametric Classification Algorithm：The tree classification procedure was modified and was proposed as a new classification algorithm which consider the overlapped region for various groups. The simulation results showed that the nonparametric classifier of the proposed method was better than the Bayesian classifier.
According to the research, we integrated two-classification systems：(I) Combined the optimal single feature selection algorithm and the order tree classifier; (II) Combined the optimal multiple features selection algorithm and the adaptive Bayesian classifier. And, the developed classification system (I) was applied to the solder joints defects of a PCB. The confidence interval of classification correctness is (98.5%, 100%), which was better than those of the optimal Bayesian classifier and the discriminate function classifier. On the application of the system II, the average classification correctness rate reached 94%, which was better than those of the Bayesian classifier, the discriminate function classifier, and the Nearest Neighbor classifier. In addition, applied both system to a set of OCR data from literature, the classification correctness rate is higher than that reported in the literature.

並列關鍵字

Automatic visual inspection ； Feature selection index ； Design of experiment ； MANOVA ； Defects classification ； Solder joints ； OCR

參考文獻

1.K. Hedengren, Methodology for Automatic Image-Based Inspection of Industrial Objects, in Advanced in Machine Vision（J. Sanz, Ed., ）, pp. 160-191, Springer-Verlag, New York, 1989.

3.T. S. Newman, and A. K. Jain, “A Survey of Automated Visual Inspection,” Computer Vision and Image Understanding, Vol. 61, No. 2, pp. 231-262, 1995.

4.H. Freeman, Machine Vision for Inspection and Measurement, Academic Press, New York, 1989,

5.A. K. Elshennawy, “The Role of Inspection in Automated Manufacturing,” Computers and Industrial Engineering, 17, pp. 327-332, 1989.

9.W. E. Eureka, and R. Nancy, Quality Up, Costs Down: A Manager''s Guide to Taguchi Methods and QFD, Dearborn, Mich.: ASI Press; Burr Ridge, Ill. :Irwin Professional Pub, 1995.