透過您的圖書館登入
IP:13.59.34.87
  • 學位論文

應用支持向量資料描述於LED瑕疵檢測

Light-Emitting Diode Defect Detection based on Support Vector Data Description

指導教授 : 劉益宏

摘要


摘要 發光二極體(Light Emitting Diode, LED),隨著環保意識的高漲,LED的量產腳步也逐漸加快,而在這波競價風暴的,為了取得最大的獲利,提高LED晶粒的良率為最主要的一個重要目標,因此,利用自動化瑕疵檢測系統可以減低人為因素並且增加檢測時間的速度,並且可以利用瑕疵檢測結果去檢查出製程中哪裡出現問題,並且加以改善,減少瑕疵問題的產生。 在本文中,主要偵測的晶粒Pad區和發光區,流程主要分為三大部分,分別為影像前處理、影像的特徵抽取和分類器的訓練。影像前處理是將影像校正,並且切割成數張大小一樣的子影像,然後對子影像進行特徵抽取的方式,抽取的特徵包括了離散餘弦轉換(Discrete Cosine Transform)、紋理特徵(Texture Feature)、影像功率(Image Power),然後將這些特徵值排列成特徵向量,用來訓練支持向量資料描述 (Support Vector Data Description, SVDD)分類器以及二值化分類進行二段偵測,使得偵測率達到95.94%,誤判率只有3.7%。

並列摘要


Abstract The Light Emitting Diode (LED) is showing rapid progress in LED’s production line when environmental protection consciousness gains ground .It is an important goal which enhancing the production yield rate of LED’s products for Raising more profit .Therefore, using automatic defects inspection system for LED can reduce human mistake and inspection time, it also can find the problem of machine to avoid LED’s defects. In this research explores the detection of light-emitted area, P-electrode and N-electrode, this system would be inspecting the defect with three mechanisms: Vision Pre-processing, Feature extraction, Training Procedure. Vision Pre-processing is made an adjustment in original defect images, and then to decompose the image to several sub-images. Feature extraction is construction of Discrete Cosine Transform, Texture, Image Power, and Statistics, according to these features, Support Vector Data Description eigenvector is trained with these features. Using to the Support Vector Data Description and the binary image classification to class with LED images.

參考文獻


[4] 郭家成,“薄膜電晶體液晶顯示器Mura瑕疵檢測技術之研發”,國立臺北科技大學自動化科技研究所碩士論文,2006
[7] 陳佳源,“結合平移距離與傅立葉特徵之手勢辨識系統”,私立中原大學資訊工程學系碩士論文,2007
[15] 拓墣產業研究所,LED技術與應用發展日趨多元化,拓墣科技,2009
[1] 洪崇祐,“應用一維傅立葉分析於TFT-LCD液晶顯示面板之瑕疵檢”,私立元智大學工業工程與管理研究所碩士論文,2004
[2] 陳興倫,“使用小波包分解與部分最小平方法於光學元件之可視瑕疵檢測”,私立朝陽科技大學工業工程與管理研究所碩士論文,2009

延伸閱讀