現今電子產業蓬勃發展,樣品多元化已是世界趨勢,企業為追求更大利益除擴大生產之外,更重要的是降低成本。而在印刷電路版的組裝測試上,自動光學檢測儀器(Automatic Optical Inspection, AOI)雖被廣泛的運用於製程中,但辨識的過程仍存有許多應被檢出卻未被檢測出的瑕疵(亦即漏失),以目前的特徵技術,往往需要足夠的樣品數做離線訓練才能達到較精確的成效,少量多樣的產品往往不適用此方式,本研究將針對這些組裝後之印刷電路板做元件的瑕疵檢測方式與分類做探討與改進。 本研究應用文獻中之幾項演算法(如:相關係數法、白點統計法…..等),並輔已實際生產線上所擷取之影像為樣本做檢測且分析各個演算法指標的優劣。由實驗結果證明單一演算法並無法完全的檢測出所有的元件是否合格,而檢測標準在於所判斷的閥值是否定義適當,且目前缺陷分類時之誤判率亦相當高,造成後段維修的困擾,因此本研究希望找到不同的演算方法來來改善缺陷影像漏失的情況,並導入倒傳遞神經網路將檢測的結果予以分類,以漏失率與分類正確率來評估與現有技術上的檢測效果。 經實驗證明,吾人所提的區域切割法可以較其它指標更能夠將缺陷影像檢測出,透過倒傳遞神經網路分類亦有較佳的成效。
The electronic industry grows vigorously now, and the pluralism of the product has already been a trend of the world. In order to pursue greater interests except expanding production, it’s more important that cost-down in enterprises. AOI (Automatic Optical Inspection) is applied in the production flow, but we don’t make sure that all of the effects could been found by current algorithms. We often need more and more sample to count and train when taking off machine. But a small amount of various products are not suitable to apply this way. I use several items of current algorithms, for example: the coefficient correlation law, white point statistic law…etc.), and compare the images from on-line process to calculate these indicators of each algorithm for analysis the sample actually. I hope to find other different methods of mathematical calculations to improve the defect images. We both know that we could not find all defects by a single algorithm, so I use Back propagation Network for sorting out what kinds of defect it is. According to my experiment report, I’m sure I can improve the missing situation, and also have a better result in sorting out.