博碩士論文查詢聯邦系統首頁 聯絡信箱 關於我們 加入我們

本論文已被瀏覽 49 次, [ 造訪詳細資料與全文 ] 9 次,[ 回到前頁查詢結果 ] [ 重新搜尋 ]

應用倒傳遞類神經網路於自動光學檢測機之解碼器輔助圖像辨識系統

Using Back-Propagation of Artificial Neural Network for Decoder Auxiliary Image Recognition System of Automatic Optical Inspection

作者:胡文瀚
畢業學校:國立高雄第一科技大學
出版單位:國立高雄第一科技大學
核准日期:2016-01-07
類型:Electronic Thesis or Dissertation
權限:Copyright information available at source archive--nkfust....

中文摘要

面板廠於製造生產過程中 , 難免會造成製程上的異常 ,而這些異常缺陷(Defect)必須透過自動光學檢測機(AOI, Automatic Optical Inspection)來進行檢測攔截,以防止產品異常區間擴大,同時可針對異常部分進行立即性處置。另外,針對異常缺陷部分可進行不同的分析,若為前製程造成的缺陷,同時有影響到外觀、電性或後製程作業時,則會將此批異常貨拉至雷射(Laser)修補機並利用雷射將其缺陷清除; 另一方面,若為當站製程造成的缺陷,則會拉回重工
但是,基板於修補機進行雷射作業前 ,還有一個關鍵要素,就是需將基板ID利用解碼器解碼出來,這樣才知道欲進行雷射作業的基板是哪一片,但ID的製程來源多樣,解碼器尚未有能力將所有不同的製程ID全部解碼出來,進而演變成後續倘若無法解碼的ID ,便利用AOI機台的固定點截圖功能,再安排人員開啟圖片利用目視方式判讀ID 。這樣的方式不但耗費時間、人力及成本,甚至延遲產品的出貨時間。因此,本文中提出應用倒傳遞類神經網路 ( BPN , Back Propagation Neural Network )之網路架構,並利用AOI機台的固定點截圖功能來收集所有不同製程的ID圖片,結合BPN輸入與輸出的邏輯關係,形成一個輔助解碼器的圖像辨識系統。
本實驗結果,機台進行基板ID截圖,驗證BPN網路模型,確實可學習辨識圖像,隨機測試辨識精準度可達98%,未來可嘗試其他較複雜參數調整及討論網路之收斂特性。

英文摘要

It is inevitable having abnormal defects during the production process, that must be detected and intercepted by Automatic Optical Inspection (AOI) in the panel plant this way. Based on this way, it can minimize the non-conforming product scope and take actions for the non-conformance immediately. In addition, if that defects due to pre-process and affect the appearance, the electric characteristics or the post-process. Then the non-conforming products will be repaired by Laser Repair System, and the defects will be fixed. On the other hand, it will require rework process if it is due to the current process.
However, before the laser repair stage it requires another key process decoding the substrate ID. This ID be identified the target substrate for laser repair stage. Since ID has various process sources, the decoder is still incapable of decoding the ID of all processes. For the ID that can t be decoded, it will use AOI s imaging function at fixed points, and then dispatch personnel to view the ID on the pictures. It wastes a lot of time, manpower and cost . Even delays the product delivery. Therefore, In this research the network architecture of Back-Propagation Neural Network (BPN) with AOI s imaging function at fixed points to collect the ID pictures of all processes is proposed. By integrating the logic relationship of BPN input and output, it forms a graphical identification system to assist the decoder.
The experimental results show, that the accuracy of identification of the substrate ID reaching 98% on the random testing with BPN model. In the future, it can try adjusting other complicated parameters and discussing the convergence characteristics of the network.


指導教授 - 黃勤鎰

委員 - 李青旻

召集委員 - 鄭竣安


 

計畫贊助者: