矩陣發光二極體（Matrix Light Emitting Diode, Matrix LED）是業界應用最廣泛的LED材料之一。因為矩陣LED是個低單價的產品，加上檢測機台成本太高，使得廠商購買意願降低，所以矩陣LED瑕疵缺陷仍然由人工進行檢測。隨著人工成本提升和人工檢測的不穩定性，我們需要應用自動光學檢測(Automated Optical Inspection, AOI)解決矩陣LED瑕疵檢測的問題。在這本論文研究中，我們提出了一套有效的矩陣LED檢測系統。該系統提供三種檢測，第一為表面刮傷瑕疵、第二為RGB亮暗點檢測、第三為使用支援向量機(Support Vector Machine, SVM)進行亮暗點分析。
在表面刮傷瑕疵檢測主要由SURF (Speeded-Up Robust Features)特徵匹配結合透視變換進行圖像校正，然後在輪廓檢測部分本文使用FindContours函式，接著找出瑕疵邊緣使用Canny邊緣檢測，該方法的準確度可達98.00％，檢測每顆矩陣LED需花2.95秒。
RGB亮暗點檢測使用ROI擷取每顆LED後，使用cvAvgSdv函式計算R、G、B平均值，首先與前一顆LED進行G值的比較，將有色差LED檢測出來，最後與制訂範圍進行判斷，該方法的準確度可達到98.00%，實驗結果顯示，檢測每顆矩陣LED需花0.01秒。
最後，使用SVM結合HOG進行圖像分類，解決矩陣LED亮暗點的問題，其準確率可達98.33％，執行速度上，檢測每顆矩陣LED需花0.38秒。實驗結果顯示，所提出的方法是有效的，並勝過以前的方法。

The Matrix Light Emitting Diode (Matrix LED) is one of the most widely used LED material in the industry. With a low unit price, Matrix LED manufacturers don't have much incentive to invest in expensive inspection machines, so the job of detecting defects still relies on manual work. With the rise of the labor cost and the instability of the manual inspection quality, it is needed to apply Automatic Optical Inspection (AOI). In this study, we propose an effective AOI system for Matrix LED. The system provides three defect inspections, Surface defects, RGB Bright/Dark spot, and SVM Bright/Dark spot analysis.
In this paper, Surface defects detection is mainly done through image correction by SURF feature matching and Perspective Transformation. Then we find contour with the findCoutours function and finally use the Canny edge detection to detect defect edges. This technique has a 98.00% accuracy and it takes about 2.95 seconds to examine each Matrix LED.
For RGB Bright/Dark spot detection, we use ROI to capture each LED, obtain mean RGB values with the cvAvgSdv function, see if they fall within defined ranges and compare the G value with the previous LED to locate LEDs with color differences. This technique has a 98.00% accuracy and it takes about 0.01 seconds to examine each LED Matrix.
Lastly, we combine SVM with HOG for image classification to remedy Bright/Dark spots. This procedure has a 98.33% accuracy and it takes about 0.38 seconds to examine each LED Matrix.
Experimental results show that the proposed methods are effective and outperform the previous methods.

[1]LED參閱自維基百科全書https://en.wikipedia.org/wiki/Light-emitting_diode
[2]吳易錡，“太陽能晶片表面瑕疵檢測”，聖約翰科技大學自動化及機電整合研究所碩士論文，2011
[3]黎晃銘，“有機發光二極體表面瑕疵檢測系統開發”， 國立臺灣科技大學高分子系碩士論文，2009
[4]賴嘉偉，“玻璃表面瑕疵的自動化檢測”， 聖約翰科技大學自動化及機電整合研究所碩士論文，2012
[5]Ssu-Han Chen, Hsiao-Wei Liu, Der-Bau Perng,“A LED Substrate Auto-Inspection System,” International Conference on Machine Learning and Cybernetics (ICMLC), 2016
[6]黃思博，”應用自動化光學檢測技術於TFT-LCD模組瑕疵辨識系統之開發與研究”，國立台灣科技大學自動化及控制研究所碩士學位論文，2013
[7]Chun-Chieh Liao, Shun-Pin Hsu, Feng-Chi Cheng, Hui-Ya Chow, Yuh-Chin Chang, “A Cost-Effective Image-Processing Approach to Improving the Uniformity of LED Panels,” IEEE Conference on Industrial Electronics and Applications (ICIEA), 2013 8th.
[8] C.-W. Hsu, C.-C. Chang, C.-J. Lin. A practical guide to support vector classification . Technical report, Department of Computer Science, National Taiwan University. July, 2003.
[9]莊揚擇，“TFT-LCD陣列製程瑕疵之檢測與分類”，國立勤益科技大學資訊工程系碩士論文，2014
[10]陳衍均，“應用支持向量資料描述於LED瑕疵檢測”，中原大學機械工程學系碩士學位論文，2012
[11]梁梓遠，“應用自動化光學檢系統技術於發光二極體微觀瑕疵辨識系統之開發與研究”，國立臺灣科技大學自動化及控制研究所碩士論文，2011
[12]Xian-Ren Chang, Chung-Yen Su, “Automated Matrix LED Detection System,” Proceedings of the 14th International Conference on Automation Technology, December 8-10, 2017.
[13]林裕軒、楊智翔、黃世霖、歐聰憲，“LED車燈自動化光學量測系統”，第12全國AOI論壇與展覽大會手冊，2012
[14]毛星雲、冷飛雪，“OpenCV程式設計參考手冊”，松崗出版社，2015
[15]Opencv中使用SURF特徵實現圖像配準及對透視變換矩陣H的平移修正 https://blog.csdn.net/dcrmg/article/details/52627726
[16][OpenCV3]透視變換https://blog.csdn.net/guduruyu/article/details/72518340
[17]S. Suzuki, “Topological structural analysis of digitized binary Images by border following,” Computer Vision, Graphics, and Image Processing, vol. 30, no. 1, pp. 32-46, Apr. 1985.
[18]唐唯傑，“具頭動補償功能之頭戴式眼動儀系統”，國立臺灣師範大學電機工程學系碩士論文，2016
[19]J. F. Canny. “A computational approach to edge detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence. vol. 8, no. 6, 1986, pp. 679-698.
[20]影像平滑 http://monkeycoding.com/?tag=%E5%BD%B1%E5%83%8F%E5%B9%B3%E6%BB%91
[21]OpenCV統計應用-平均數與標準差 http://yester-place.blogspot.tw/2008/10/opencv-cvavgcvavgsdv.html
[22]林宗勳，“Support Vector Machines 簡介” http://www.cmlab.csie.ntu.edu.tw/~cyy/learning/tutorials/SVM2.pdf
[23]莊楊擇，“TFT-LCD陣列製程瑕疵之檢測與分類”，國立勤益科技大學資訊工程系碩士在職專班，2013
[24]HOG基本原理完整介紹https://blog.csdn.net/gy429476195/article/details/50156813
[25]DFK 27AUP006 USB 3.0 彩色工業相機 https://www.theimagingsource.tw/%E7%94%A2%E5%93%81/%E5%B7%A5%E6%A5%AD%E7%9B%B8%E6%A9%9F/usb-3.0-%E5%BD%A9%E8%89%B2/dfk27aup006/
[26]M1614-MP2 https://computar.com/product/554/M1614-MP2
[27]SGSP Series Translation Motorized Stages 產品系列 https://www.global-optosigma.com/en_jp/Catalogs/gno/?from=page&pnoname=SGSP26-%28XY%29&ccode=W9015&dcode=&gnoname=SGSP26-150%28XY%29
[28]Single axis stage controller GSC-01 https://www.global-optosigma.com/en/page_pdf/GSC-01.pdf
[29]KT-89S51 V3.3燒錄板 http://www.ltc.com.tw/product_info.php/products_id/59987