透過您的圖書館登入
IP:18.206.160.129
  • 學位論文

運用卷積神經網路建立積體電路封裝缺陷分類檢測模型

Development of Defect Classification and Detection Models for IC Packaging Using Convolutional Neural Networks

指導教授 : 項衛中

摘要


在現今科技產品的廣泛運用下,相關電子產業蓬勃發展,半導體晶片封裝的研發也朝向高效能與輕量化,以滿足電子產品的需求。利用機器視覺與神經網路分類的方式來辨別半導體封裝晶片缺陷與種類,將可大量降低人工檢驗產品缺陷的成本,並提升檢測速度和準確率。本研究運用卷積神經網路與Mask R-CNN兩種演算法,及不同分類種類與晶片影像共三種因子,建立探討晶片封裝的缺陷分類與檢測模型,進而探討各因子對模型的影響度。 本研究所建立的檢測模型可分為四個部分,第一部分為影像資料前處理,將蒐集到的影像資料切割成單一晶片;第二部分為影像資料擴增處理,將影像數量過少的缺陷類別,提取缺陷特徵後複製在良品影像上,使良品與不良品資料數量達到平衡;第三部分為訓練資料的前處理,將影像資料整理成演算法可判讀的格式;第四部份為模型訓練與驗證,運用實驗設計,分析實驗因子對分類結果的影響。研究結果發現Mask R-CNN所建立的模型比卷積神經網路所建立的模型更能在較複雜的影像中得到較準確的分類結果,同時因Mask R-CNN的標註特性,判斷缺陷時能顯示出缺陷位置,能夠得到更完整的預測結果。此外透過實驗結果也發現到若分類種類分得越多,則模型的判斷準確度也會跟著下降;晶片影像結構較為簡單的影像,也能得到較準確的檢測結果。

並列摘要


With the widespread use of high-tech products, the related electronic industry is booming, therefore, the development of semiconductor chip packaging is moving toward high performance and light weight to meet the needs of electronic products. Using machine vision and neural network to identify defects and their types of semiconductor package has significantly reduced the manual inspection cost for product defects and improved the inspection speed and accuracy. In this study, two algorithms, convolutional neural network (CNN) and Mask Region-based Convolutional Neural Networks (Mask R-CNN), are used to develop a defect classification and detection model for IC packaging, and then other factors: classification type and chip image are included to investigate the effect of each factor on the model. There are four stages for developing the classification model. The first stage is pre-processing image data, which cuts the whole image data into a single chip. The second stage is augmenting defect image data, which extracts defect parts from the defect images and copies them on the normal images. So that the number of defect image would increase and could be balanced with the normal images. The third stage is transforming pre-processed images data into a format that can be interpreted by the algorithms. The fourth stage is model training and validation, experimental design is used to analyze factor effects on classification accuracy. It was found that the Mask R-CNN model could get more accurate classification results in more complex images than convolutional neural network model, and the labeling feature of Mask R-CNN showed the defect location when it detected the defects. In such a case, Mask R-CNN generated more complete prediction results. In addition, it was found that the accuracy of the model decreased in more defect categories cases, and for simpler chip patterns, the models may have better classification accuracy.

參考文獻


[1] David Silver, Aja Huang , Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel Demis Hassabis(2016), Mastering the game of Go with deep neural networks and tree search, Nature vol 529.
[2] 上報(2018)。中美貿易戰顯台灣「半導體王國」重要地位。民國110年11月01日,取自:https://www.upmedia.mg/news_info.php?SerialNo=45923
[3] Yimou Lee, Norihiko Shirouzu David Lague (2021), Taiwan chip industry emerges as battlefront on U.S.-China showdown, Reuters. https://www.reuters.com/investigates/special-report/taiwan-china-chips/
[4] Pat Hudson(2014), The Industrial Revolution, Bloomsbury Publishing.
[5] 工業技術研究院(2020,10月22日)。智慧製造是什麼?工業4.0下,智慧製造發展的4要素與4大挑戰。民國110年11月10日,取自:https://college.itri.org.tw/event/525-smart-machinery2020.html

延伸閱讀