本研究係發展一套發光二極體(light emitting diode, LED)自動光學檢測(automatic optical inspection, AOI)系統，應用機器視覺與倒傳遞類神經網路(back-propagation neural network)檢測LED晶粒的外觀瑕疵，檢驗的項目包括破晶、發光區遺失、發光區污染、蝕刻線不全、電極區遺失、針痕過大、電極區污染等七項。首先，在樣本擷取方面，應用相關係數法(normalize cross correlation)，以及中央加權中值法(center weight median filter)來進行檢測樣本影像的擷取與前處理。在樣本影像瑕疵檢測部份，針對晶片外型，利用統計式門檻值(Otsu thresholding)二值化、影像型態學(morphology)、索貝爾測邊運算子(Sobel operator)以及細線化(thinning)，將晶粒外型的面積與周長擷取出來，進而計算出晶粒的形狀飽和度(compactness)。針對發光區的部份，運用改良式統計式門檻值(a fast algorithm for Otsu multilevel thresholding)四值化，並搭配相對應的電極與蝕刻線遮罩，擷取出發光區瑕疵的瑕疵區域，最後統計其面積。在蝕刻線的部份，利用改良式統計式門檻值四值化與細線化，擷取出蝕刻線的特徵影像，再計算此特徵影像的一階矩以及進行充填後的面積。最後在電極區的部份，利用相關係數法，擷取出電極區影像後，對此影像使用乘冪律轉換(powerlaw transform)進行影像強化，再利用統計式門檻值、影像形態學，擷取出特徵結構影像，再計算其面積與不變矩(invariant moment)。
利用上述所得共30個特徵值，使用倒傳遞類神經網路對743個訓練樣本加以學習訓練，以分類出LED晶粒是否帶有瑕疵。最後以371個晶粒測試樣本影像檢驗系統的辨識能力，總辨識率達91.98%，證明本研究發展出一套適合於檢測LED瑕疵的自動化光學檢測系統。

This study proposes an Automatic Optical Inspection (AOI) System that applies the Machine Vision technology and Back-propagation Neural Network to inspect defects of the light emitting diode (LED). The defects include chip crack, lighting area missing, lighting area contamination, lack of etch line, pad missing, needle mark error, and pad contamination. We first applied the normalized cross correlation to obtain the single chip image in the LED sample images, and used center weighted median filter to reduce noise. For the chip contour, we applied the Otsu image morphology technique, Sobel operator and thinning to get the area and circumference of the chip. Then we used the area and the circumference to calculate the compactness for the eigen value. For the lighting area, we applied the fast algorithm for Otsu multilevel thresholding, pad filter and etch line filter. We could take out the lighting area missing and the lighting area contamination error pixels, and count this area for the eigen value. For etch line, we also used the fast algorithm for Otsu multilevel thresholding and thinning to get the etch line pixels. We calculated the area and the first order moment for the eigen value. Regarding the Pad defect, we used the normalized cross correction to take out the pad image and used the powerlaw to improve the image. We then applied Otsu thresholding and morphology to obtain the eigen area. Calculating area and invariant moment allowed us to recognize these defects of pad area.
Lastly, we use all of these eigen values to the Back-propagation Neural Network for training and study, and we recognized all kinds of defects of the LED chip. We experimented with 371 chips of LED samples, obtaining the recognize rate of 91.98%, indicating that this system is work.

1. Du-Ming Tsai and Chien-Ta Lin, “Fast normalized cross correlation for defect detection,” Pattern Recognition Letters, Volume 24, Issue 15, November 2003, Page(s):2625-2631.
2. Du-Ming Tsai, Chien-Ta Lin, and Jeng-Fung Chen, “ The evaluation of normalized cross correlations for defect detection,” Pattern Recognition Letters, Volume 24, Issue 15, November 2003, Page(s):2525-2535.
3. D. R. K. Brownigg, “The weighted median filter,” Communications of the ACM, Volume 27, Issue 8, August 1984, Page(s): 807–818.
4. Ping-Sung Liao, Tse-Sheng Chen and Pau-Choo Chung, “A fast algorithm for multilevel thresholding,” Journal of Information Science and Engineering, Volume 17, Number 5, September 2001, Page(s):713-727.
5. Ming-Kuei Hu, “Visual pattern recognition by moment invariants,” Information Theory, Volume 8, Issue 2, Feb 1962, Page(s):179-187.
6. Rumelhart, D. E., Hinton, G. E., and Williams, R. J., “Learning internal representations by error propagation,” Parallel distributed processing: explorations in the microstructure of cognition, vol. 1: foundations , Page(s): 318-362.
7. Henry A. Rowley, Shumeet Baluja, and Takeo Kanade, “Rotation invariant neural network-based face detection,” Digital Object Identifier, June 1998 Page(s):963-963.
8. Chuan-Yu Chang, Yung-Chi Chang, Chun-His Li, and Mu-Der Jeng, “Radial basis function neural networks for LED wafer defect inspection,” Digital Object Identifier IEEE, Sept. 2007 Page(s):229-229.
9. Chuan-Yu Chang, Chin-Huang Chang, Chun-Hsi Li, Mu-Der Jeng, “Learning vector quantization neural networks for LED wafer defect inspection,” Digital Object Identifier IEEE, Sept. 2007 Page(s):230-230.
10. 劉明進, “高亮度LED外觀瑕疵檢測技術,” 暨南國際大學, 2005.
11. 吳宙遠, “應用影像處理技術於高亮度LED外觀瑕疵檢測,” 臺灣科技大學, 2007.
12. 繆紹綱編譯, “數位影像處理,” 普林斯頓國際有限公司, 2004.
13. 鍾國亮編著, “數位處理與電腦視覺,” 東華書局, 2006.
14. 吳上立 林宏敦, “C語言數位影像處理,” 全華科技圖書股份有限公司, 2006.
15. 徐明興等編著, “Borland C++ Builder 6 程式設計經典,” 文魁資訊, 2002.
16. 蔡孟凱等編著, “C++ Builder 6 完全攻略,” 金禾資訊, 2005.
17.王進德, “類神經網路與模糊控制理論 入門與應用,” 全華科技圖書股份有限公司, 2007.
18. 陳隆建編著, “發光二極體原理與製造,” 全華科技圖書股份有限公司, 2006.