固晶是LED封裝測試製程中八階段的第一階段。後面七階段開始前，須先經由第一站的固晶前目檢做把關。固晶前目檢為作業人員以光學顯微鏡檢查晶片並淘汰挑出不良品， 以避免後面的製程白費功夫。由於LED封裝測試的目檢站是高勞力的作業，造成市場目檢人員的薪資水準也很高，工廠對此負擔的人力成本非常的大。因此，本研究欲解決的問題為在固晶前的目檢階段，使用影像辨識自動地檢測晶片瑕疵並取代目前的人力。

本研究先對原始影像資料做了智慧型的轉換，也就是將輸入的影像從RGB色彩空間轉換成HSV色彩空間。再經由一系列形態學的轉換，最後藉由機率密度函數找出飽和度閥值，讓晶片影像能夠很精確地被截取出來。此智慧型轉換的方法比傳統直接使用顏色或灰階的方式擷取出的影像來得完整。而要確定圖像是否為有缺陷的產品，首先將轉換過後的影像再轉成灰階影像，再利用灰階閥值轉二值化後，計算屬於瑕疵的黑色像素個數。黑色像素個數若大於或等於1 則視為有瑕疵，判斷為不良品。

研究結果顯示，每張晶片影像判斷時間從人員目檢的5秒，而改成電腦影像判斷只需要0.2秒，快了25倍。人力成本方面，以6名目檢人員計算，需支付總月薪約新台幣23.7萬元，若改成電腦影像判斷則需要一名攝影機台操作人員，只需支付月薪約新台幣3至4萬元，人力成本相差近6至8倍。因此，本研究將可有效提高產能，減少人力成本並消除作業人員在固晶前目檢的勞力負擔。

The LED package test process can be divided into eight stages. Before the start of the next seven stages, it is necessary to check the quality of LED chip before the first station, and the operator will use the optical microscope to check and eliminate the defective chips, so that the subsequent processes will not be in vain. The level of production capacity is positively related to the demand of the visual inspection personnel. Since the visual inspection station for LED packaging testing is a high-labor operation, the market salary level is also high, and the labor cost for the factory is very large. Therefore, the problem to be solved in this study is to use image recognition to automatically detect chip defects and replace the current manpower on the visual inspection stage before die bonding process.

This study makes a sensible conversion of the original image data, by converting the input image into the HSV color space, then through a series of morphological transformations, and finally finding the saturation threshold through the probability density function. The chip image can be cut out very accurately, complete with the traditional image taken directly by color or grayscale.And to determine if the image is a defective product,first convert the converted image into a grayscale image, and then use the grayscale threshold to binarize, then calculate the number of black pixels belonging to the defect. If the number of black pixels is greater than or equal to 1, it is considered to be defective, and it is judged to be defective.

The result shows that the image judgment time of each chip is 5 seconds from the visual inspection by the personnel, and it takes only 0.2 seconds to change to the computer image judgment. In terms of cost, if a visual inspection of personnel is used, the total monthly salary of 6 visual inspection personnel is calculated to be approximately TWD 237,000, and the computer image judgment requires only one camera operator to pay a monthly salary of TWD 30,000 to 40,000,and the labor cost is nearly 6 to 8 times different. Therefore, this research will effectively improve production capacity and reduce labor costs and eliminate the burden on operators before die boning process.

[1] 黃至伶. (2008). 發光二極體目檢作業人因危害評估及負荷推估.高雄醫學大學職業安全衛生研究所學位論文 , 1-244.
[2] Chang, C. C. (2011). 應用邊緣偵測和二值化影像直方圖於晶片瑕疵檢測.
交通大學電機與控制工程系所學位論文 , 1-53.
[3] 張元碩, & 彭德保. (2008). 晶粒表面缺陷自動視覺檢測系統之設計與開發
(Doctoral dissertation).
[4] 李少芳, & 陳德禮. (2007). 數字圖像旋轉實現的探討.計算機與現代化 , 2007(9), 94-96
[5] Perona, P., & Malik, J. (1991). Detecting and localizing edges composed of steps, peaks and roofs.
[6] Haralick, R. M., Sternberg, S. R., & Zhuang, X. (1987). Image analysis using mathematical morphology. IEEE transactions on pattern analysis and machine intelligence, (4), 532-550.
[7] 李瑞陽, & 莊佳文. (2006). 運用影像分塊方法於高解析衛星影像土地利用判釋精度之研 究. Journal of Photogrammetry and Remote Sensing, 11(4), 403-415.
[8] 丁曉炯, & 程大章. (2009). 基於邊緣檢測和混合高斯模型的運動目標檢測算法 (Doctoral dissertation).
[9] 李宇峻 (2011). 使用邊緣偵測輔助低對比瑕疵檢測 . (Master).
[10] Baggio, D. L. (2012). Mastering OpenCV with practical computer vision projects. Packt Publishing Ltd.