基於人工智慧之PCB瑕疵檢測技術開發

以作者查詢圖書館館藏

、以作者查詢臺灣博碩士

、以作者查詢全國書目

、勘誤回報

、線上人數：63

、訪客IP：3.147.42.197

姓名

黃柏盛(BOR-SHENG HUANG) 查詢紙本館藏

畢業系所

光機電工程研究所

論文名稱

基於人工智慧之PCB瑕疵檢測技術開發
(Development of PCB Defect Detection Technology Based on Artificial Intelligence)

相關論文

★ MOCVD晶圓表面溫度即時量測系統之開發	★ MOCVD晶圓關鍵參數即時量測系統開發
★ 應用螢光顯微技術強化RDL線路檢測系統	★ 基於 YOLO 物件辨識技術之 PCB 多類型瑕疵檢測模型開發
★ 全場相位式表面電漿共振技術	★ 波長調制外差式光柵干涉儀之研究
★ 攝像模組之影像品質評價系統	★ 雷射修整之高速檢測-於修整TFT-LCD SHORTING BAR電路上之應用
★ 光強差動式表面電漿共振感測術之研究	★ 準共光程外差光柵干涉術之研究
★ 波長調制外差散斑干涉術之研究	★ 全場相位式表面電漿共振生醫感測器
★ 利用Pigtailed Laser Diode 光學讀寫頭在角度與位移量測之研究	★ 複合式長行程精密定位平台之研究
★ 紅外波段分光之全像集光器應用	★ 太陽光譜分光器之設計

檔案

[Endnote RIS 格式]

[Bibtex 格式]

[相關文章]

[文章引用]

[完整記錄]

[館藏目錄]

至系統瀏覽論文 ( 永不開放)

摘要(中)

隨著硬體技術的發展，電路板的需求量逐年大幅增加，產業對於電路板製造品質與產品良率的要求也是與日俱增；深度學習技術也隨著硬體技術的進步而蓬勃發展。尤其是在過去的幾年內，硬體效能的快速成長讓深度學習技術得以突飛猛進，深度學習技術經常被應用在生活和工業中，比如：GOOGLE翻譯、停車場車牌識別，都是深度學習在生活中的應用。
目前電路板印刷相關產業的PCB電路板良率檢測，主要依賴自動化光學檢測系統(AOI)和人工檢測，由於AOI系統經常會出現缺陷判斷錯誤，因此需要大量人工進行檢測，使得成本提高。本研究提出了一種基於深度學習的檢測技術，檢測PCB板上的缺陷，主要目的為過濾掉AOI系統標記的「假缺陷」。經過本研究開發的深度學習神經網路系統過濾後，可以大幅減少工作量和人力成本，顯著提高PCB缺陷檢測的效率和良品率。
本次研究分為神經網路系統主體架構和數據庫主體，系統主體架構為YOLO。在AI技術發達的時代，誕生了很多物件偵測技術，經過多方考慮和篩選，我們選擇YOLO作為主要架構。YOLO是近幾年非常強大的物件偵測神經網路架構系統，廣泛應用於學術界和工業界，經過訓練，可以準確地判斷和標記目標位置。資料集部分為，廠商提供之產線AOI系統輸出之影像資料，其中包含AOI檢測系統所誤判的11種瑕疵與非瑕疵的影像資料。
本研究與其他實驗難易度不同，本研究的測試資料為廠商每個月送來當下PCB產線AOI系統輸出的影像，資料變化多端且具有一定難度，時常會遇到模型從未訓練過的資料特徵型態。相較於其他實驗，測試資料為封閉資料且資料特性相近，通常由資料庫中分割一部份成為測試資料，兩者難度差距極大，因此本研究運用更多方式分析以及調整，達到我們所設定的實驗目標。

摘要(英)

With the development of hardware technology, the demand for circuit boards has increased dramatically year by year, and the industry′s demand for quality and yield of circuit boards has also been increasing. In particular, the rapid growth in hardware performance over the past few years has allowed deep learning technology to advance by leaps and bounds. Deep learning technology is often used in everyday life and industry, for example, GOOGLE translation and car park license plate recognition.
At present, PCB yield inspection in the PCB printing-related industry mainly relies on automated optical inspection systems (AOI) and manual inspection, which requires a large amount of manual inspection due to the frequent error in defect judgement in AOI systems, resulting in higher costs.
This study proposes a deep learning-based inspection technique to detect defects on PCBs, with the main objective of filtering out ′false defects′ marked by AOI systems. After filtering by the deep learning neural network system developed in this study, the workload and labour cost can be significantly reduced and the efficiency and yield of PCB defect detection can be significantly improved.
In this study, the main architecture of the neural network system is YOLO, which is a very powerful neural network system for object detection in recent years, widely used in academia and industry. It has been trained to accurately determine and mark the location of targets.
The dataset consists of images from the AOI system of the production line provided by the manufacturer, which contains images of 11 types of defects and non-defects that were misidentified by the AOI inspection system.This study is different from other experiments in terms of difficulty.
The test data in this study is the current image output from the AOI system of PCB production line sent by the vendor every month. Compared to other experiments, the test data is closed data with similar characteristics, and usually a part of the database is divided into test data, which is extremely difficult.

關鍵字(中)

★ YOLO
★ 深度學習
★ CNN
★ PCB
★ AOI

關鍵字(英)

★ YOLO
★ Deep learning
★ CNN
★ PCB
★ AOI

論文目次

摘要 I
Abstract II
致謝 IV
圖目錄 IX
表目錄 XIII
第一章緒論 1
1-1研究背景 1
1-2文獻回顧 2
1-2-1 CNN 2
1-2-2 YOLO 7
1-3 研究動機、目的與方法 9
1-4 論文架構 10
第二章人工智慧與神經網路原理介紹 12
2-1人工智慧 12
2-1-1弱人工智慧 12
2-1-2強人工智慧 12
2-2機器學習 13
2-2-1監督式學習 13
2-2-2半監督式學習 13
2-2-3非監督式學習 13
2-2-4神經網路 13
2-2-5卷積神經網路 14
2-2-6 YOLO 物件偵測 16
2-3小結 18
第三章系統架構 19
3-1基本構想 19
3-2使用設備 20
3-3前期實驗研究可行性評估結果 21
3-3-1訓練結果 21
3-3-2 CNN網路遇到困難 23
3-3-2前期實驗研究問題與檢討 23
3-3-3物件追蹤模型實驗可行性評估 25
3-4資料集 25
3-4-1 訓練集 32
3-4-2 驗證集 32
3-4-3 測試集 32
3-5標註 33
3-5-1標註軟體 33
3-5-2標註軟體操作方式 33
3-5-3標註檔 35
3-6神經網路架構 35
3-7小結 36
第四章研究方法 37
4-1實驗流程 37
4-2圖片標註 37
4-2-1短路 38
4-2-4突起 41
4-2-5線路壓傷 43
4-2-6斷路 46
4-3模型訓練 49
4-4模型評估方式：二元分類指標 49
4-5分類標準規範 48
4-6小結 52
第五章實驗設計、方法與結果 53
5-1學習率 53
5-2增加訓練資料 54
5-3輸入圖片解析度調整 55
5-4標註方式修正 56
5-5新增類別 57
5-6數據增強 58
5-6-1上下左右翻轉 59
5-6-2馬賽克 60
5-6-4 HSV調整 61
5-6-5縮放 62
5-6-6旋轉 62
5-7正則化 63
5-8人工假圖製作 66
5-9資料集比例分析 68
第六章問題討論 69
6-1問題與討論 69
6-1-1各項模型指標 69
6-1-2神經網路架構 70
6-1-3資料庫 71
6-1-4資料標註 71
第七章結論與未來展望 73
7-1結論 73
7-2 未來展望 74
7-2-1準確率提升 74
7-2-3優化辨識速度 77
7-2-4 AOI機台串聯 78
7-2-5 GUI介面增加 78
參考文獻 79

參考文獻

[1] Venkat Anil Adibhatla, Jiann-Shing Shieh, Maysam Abbod, Huan-Chuang Chih, “Detecting Defects in PCB using Deep Learning via Convolution Neural Networks,” 2018 13th International Microsystems, Packaging, Assembly and Circuits Technology Conference (IMPACT), pp. 202-205, (2018).
[2] Barnajit Ghosh, Manas Kamal Bhuyan, “Defect Classification of Printed Circuit Boards based on Transfer Learning,” 2018 IEEE Applied Signal Processing Conference (ASPCON), DOI:10.1109/ASPCON.2018.8748670, (2018).
[3] ImageNet.
https://www.image-net.org/index.php
[4] Yu-Shan Deng, An-Chun Luo, Min-Ji Dai, “Building an Automatic Defect Verification System Using Deep Neural Network for PCB Defect Classification,” Digital PCB Design & Manufacturing Department Industrial Technology Research Institute, (2018).
[5] Sahand Abbasi, Sahand, Haniyeh Abdi, Ali Ahmadi, “A Face-Mask Detection Approach based on YOLO Applied for a New Collected Dataset,” 2021 26th International Computer Conference, Computer Society of Iran (CSICC) (2021): 1-6, (2021).
[6] 人體神經元.
https://1.bp.blogspot.com/-HuMpH-XL_xU/VGti6RNknHI/AAAAAAAAH8w/JGas8MHu6l0/s1600/Screenshot%2Bfrom%2B2014-11-18%2B23%3A16%3A42.png
[7] 神經網路.
https://www.tibco.com/sites/tibco/files/media_entity/2021-05/neutral-network- diagram.svg
[8] 卷積基本運算.
https://medium.com/jameslearningnote/%E8%B3%87%E6%96%99%E5%88%86%E6%9E%90-%E6%A9%9F%E5%99%A8%E5%AD%B8%E7%BF%92-%E7%AC%AC5-1%E8%AC%9B-%E5%8D%B7%E7%A9%8D%E7%A5%9E%E7%B6%93%E7%B6%B2%E7%B5%A1%E4%BB%8B%E7%B4%B9-convolutional-neural-network-4f8249d65d4f
[9] Yann LeCun, Léon Bottou, Yoshua Bengio, Patrick Haffner, “Gradient-based learning applied to document recognition,” Proc. IEEE 86 (1998): 2278-2324. (1998).
[10] Yann LeCun, Yoshua Bengio, “Convolutional networks for images, speech, and time series,” The handbook of brain theory and neural networks. MIT Press, Cambridge, MA, USA, 255–258, (1998).
[11] Yann LeCun, Bernhard Boser, John Denker, Donnie Henderson, R. Howard, Wayne Hubbard, Lawrence Jackel, “Handwritten digit recognition with a back-propagation network,” MIT Press, Cambridge, MA, USA, 396–404, (1989).
[12] 卷積提取邊界特徵.
https://ithelp.ithome.com.tw/upload/images/20220523/20138527zd9CVR2Get.png
[13] 基本卷積神經網路架構.
https://chtseng.wordpress.com/2017/09/12/%E5%88%9D%E6%8E%A2%E5%8D%B7%E7%A9%8D%E7%A5%9E%E7%B6%93%E7%B6%B2%E8%B7%AF/
[14] Grad-CAM.
https://github.com/yizt/Grad-CAM.pytorch
[15] Kaiming He, Georgia Gkioxari, Piotr Dollár Ross Girshick, “Mask R-CNN,” 2017 IEEE International Conference on Computer Vision (ICCV), pp. 2980-2988, doi: 10.1109/ICCV.2017.322, (2017).
[16] Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, pp. 779-788, (2016).
[17] Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra Malik, “Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation, ” 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 580-587, doi: 10.1109/CVPR.2014.81, (2014).
[18] Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” in IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 39, no. 06, pp. 1137-1149, (2017).
[19] Wang Chien-Yao, Bochkovskiy Alexey, Liao Hong-yuan. “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” 10.48550/arXiv.2207.02696, (2022).
[20] Liao Xinting, Shengping Lv, Denghui Li, Yong Luo, Zichun Zhu, Cheng Jiang, “YOLOv4-MN3 for PCB Surface Defect Detection,” Applied Sciences 11, no. 24: 11701. https://doi.org/10.3390/app112411701, (2021).
[21] YOLO人臉辨識成果.
https://www.youtube.com/watch?v=rzjGz20WUyM
[22] Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton. “ImageNet classification with deep convolutional neural networks,” In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1 (NIPS′12). Curran Associates Inc., Red Hook, NY, USA, 1097–1105, (2012).
[23] Yufeng Li, Shengli Lu, Jihe Luo, Wei Pang, Hao Liu, “High-performance Convolutional Neural Network Accelerator Based on Systolic Arrays and Quantization,” 2019 IEEE 4th International Conference on Signal and Image Processing (ICSIP), pp. 335-339, doi: 10.1109/SIPROCESS.2019.8868327, (2019).
[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, “Deep Residual Learning for Image Recognition,” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770-778, doi: 10.1109/CVPR.2016.90, (2016).
[25] Ramprasaath Ramasamy Selvaraju, Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, Dhruv Batra “Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization,” 2017 IEEE International Conference on Computer Vision (ICCV), pp. 618-626, doi: 10.1109/ICCV.2017.74, (2017).
[26] Labelimg.
https://github.com/heartexlabs/labelImg
[27] 陳雲，“比Tensorflow還精美的人工智慧套件Pytorch讓你愛不釋手,”，佳魁數位, (2018).
[28] YOLO V5.
https://github.com/ultralytics/yolov5
[29] Coco dataset.
https://cocodataset.org/#home
[30] Learning rate 數值大小對梯度下降的影響.
https://ithelp.ithome.com.tw/articles/10204032
[31] Mosaic.
https://blog.csdn.net/qq_41011242/article/details/110439183
[32] HSV.
https://zhuanlan.zhihu.com/p/67930839
[33] Underfittng.
https://miro.medium.com/max/1200/1*UCd6KrmBxpzUpWt3bnoKEA.png
[34] 葉欣睿，“Deep learning 深度學習必讀：Keras 大神帶你用 Python 實作,” 旗標出版, (2019).
[35] 李金洪，“詳細+超深入：最新版TensorFlow 1.x/2.x完整工程實作,” 深智數位, (2020).
[36] Kurt Smith, “Cython,” O’Reilly Media, (2015).
[37] Tesla V100 v.s Tesla A100 Masked-R-CNN運行速度.
https://zhuanlan.zhihu.com/p/355146807

指導教授

李朱育(Ju-Yi Lee)

審核日期

2023-2-1

推文