臺灣博碩士論文加值系統

隨著人工智慧（Artificial Intelligence；AI）興起與發展，主導著未來趨勢，生活中到處可見電子產品，然而內部零件之焊接處，會影響著產品品質優劣，因此檢驗就顯著相當重要，這階段檢驗還是避免不了人工目視檢查，針對焊接點察看或是有接觸不良、錫橋、立碑、缺件、歪斜和損毀等情況，長時間檢查下來會容易造成眼睛負擔，導致誤判情況。
為了改善上面所發生的問題，本研究透過深度學習(Deep Learning)對資料與圖像進特徵學習，利用建立網路架構與演算法計算，使機器之判斷與思維接近我們預設之目標，幫助作業流程檢測與製程更加快速及便利。以OpenR8軟體協助完成此項實驗，採用卷積神經網路（Convolutional Neural Network；CNN）為原理，利用Caffe建立與軟體溝通橋梁，針對PCB電子零件影像進行瑕疵檢驗，透過SSD(Single Shot MultiBox Detector)架構使機器訓練更有效率與精準，並分析瑕疵種類幫助品檢人員在生產線上進行篩檢，有效降低漏失率(Underkill rate)與誤判率(Overkill rate)以達到本研究之目的，後續也可以透過瑕疵品進行製程能力分析找出相關問題並克服與改善。
依照實驗結果及實務測試，兩種演算比較之下， SSD學習框架明顯在各方面評估指標都優於ResNet演算法，此模型準確率高達97.3%誤判率僅2.7%，且並無產生其漏失情況，僅有部分少辨識及誤判之結果，後續可透過學習計畫將誤判特徵加入下次訓練中以增加模型記憶，修正其模型辨識能力，導入AI之優勢便可依照其AI分析師與品管人員進行分析決策，透過生產記錄瑕疵回傳等系統流程，逐漸改善修正，提升整體競爭力。

With the rise and development of Artificial Intelligence (AI), which dominates the future trend, electronic products can be seen everywhere in life. However, the soldering of components will affect the quality of the products. Therefore, the inspection is obviously very important. It is still impossible to avoid manual visual inspection. For the inspection of soldering points, there are cases of poor contacts,bridging, tombstone, missing component, shift and damage. If you check for a long time, it will easily cause eye burden, leading to misjudgment.
This study uses Deep Learning to learn the characteristics of data and images, and uses network architecture and algorithm calculations to make the judgment and thinking of the machine close to our preset goals, helping the process detection and process faster and more convenient. OpenR8 is adopted to assist in the completion of this experiment, a Convolutional Neural Network (CNN) is as a deep learning algorithm, using Caffe to establish a bridge with software, for PCB electronic parts image inspection, through SSD (Single Shot MultiBox Detector). The structure makes the machine training more efficient and accurate, and analyzes the types of defects to help quality inspectors perform screening on the production line. It effectively reduced the Underkill rate and Overkill rate to achieve the purpose of this research, and then can also find the process capability analysis through the products related issues are improved.
According to the experimental results and practical tests, under the two kinds of calculations, the SSD learning framework is obviously superior to the ResNet algorithm in all aspects. The accuracy rate of this model is as high as 97.3%, the false positive rate is only 2.7%, and there is no leakage. Only part of the results of less recognition and misjudgment can be added to the next training through the learning plan to increase the model memory, correct its model identification ability, and introduce the advantages of AI according to its AI analyst and quality control. Personnel conduct analysis and decision-making, and gradually improve the revision and improve the overall competitiveness through system processes such as production records and defect return.

摘要 i
Abstract ii
目錄 iv
圖目錄 vi
表目錄 vii
第一章緒論 1
1.1研究動機 1
1.2研究目的 1
1.3研究範圍與限制 2
1.4論文研究 2
第二章文獻回顧 4
2.1人工智慧演進 4
2.2機器學習 5
2.3深度學習 6
2.4CNN模型 9
2.4.1LeNet模型 10
2.4.2AlexNet模型 10
2.4.3VGGNet模型 11
2.4.4GoogleNet模型 12
2.5程式語言及框架 13
2.5.1 Python 14
2.5.2 C++ 14
2.5.3 Caffe 15
第三章研究方法 16
3.1電子元件 16
3.2研究架構 18
3.3軟硬體規格 19
3.4實驗數據說明 20
3.5實驗網路架構及演算法 20
第四章實驗過程 23
4.1實驗架構 23
4.2採樣及訓練 25
4.3數據呈現與比較 28
4.3.1 SSD512×512 28
4.3.1 ResNet512×512 34
4.4實驗結果 40
4.4.1 SSD實務測試 40
4.4.2 ResNet實務測試 44
第五章結論 52
參考文獻 54
中文文獻 54
英文文獻 55
附錄 57

中文文獻
[1]李建興，2018，IEEE Spectrum發布程式語言排名：Python持續強勢、R語言逐漸消退，上網日期：107年11月21日，檢自：https://www.ithome.com.tw/news/125008
[2]邱昰芳，2018，兩岸PCB產業競合與台商因應之道 ，兩岸經貿第319期。
[3]林大貴著，2017，TensorFlow＋Keras深度學習人工智慧實務應用，博碩文化，台灣。
[4]林宜臻，2018，基於深度學習之戶外導航機器人，國立中央大學電機工程學系，碩士論文。
[5]洪文麟，2016，深度學習應用於以影像辨識為基礎的個人化推薦系統-以服飾樣式為例，國立成功大學工程科學所，碩士論文。
[6]胡程維，2017，程式知識課（一）。上網日期：107年12月10日，檢自：https://medium.com/appworks-school/most-popular-programming-language-for-first-time-learners-54b7da125e3c
[7]許明翔，2015，具有深度學習精神之人類手勢影像辨識系統，國立臺北科技大學自動化科技研究所，碩士論文。
[8]郭冠毅，2018，植基於SSD之孿生網路應用於魚追蹤之研究，國立臺南大學資訊工程學系，碩士論文。
[9]陳光軒，2018，利用深度學習單次多框檢測器在快速道路上進行車輛計數，臺北市立大學資訊科學系，碩士論文。
[10]陳怡霖，2018，基於深度學習目標檢測之羽球影片空間定位及球員位置預測，國立交通大學資訊科學與工程研究所，碩士論文。
[11]張耀仁，2018，改進單次多框偵測器架構與後處理，國立臺灣大學電信工程學研究所，碩士論文。
[12]傅汝鴻，2018，基於深度學習之天候影像分類，國立雲林科技大學資訊工程系，碩士論文。
[13]蔡典霖、蘇家興、林士傑，2004，應用類神經網路於印刷電路板表面元件之機械視覺檢測，第四屆精密機械製造研討會，台北，11月13日。
[14]黎榮富，2017，以DOE手法降低SMT光學檢測設備檢測誤判率，國立臺北科技大學工業工程與管理系，碩士論文。
[15]蘇家興，2005，印刷電路板表面黏著元件視覺檢測系統，國立清華大學動力機械工程所，碩士論文。
[16]KDnuggets，2017，谷歌趨勢：人工智慧首超大數據，深度學習關注度上升最快，上網日期：107年11月21日，檢自：https://zi.media/@yidianzixun/post/8szx9F
[17]每日頭條：深度學習｜高階機器學習庫-Caffe，上網日期：107年12月11日，檢自：https://kknews.cc/tech/oynymxq.html
[18]電子元件-Wikiwand，上網日期：107年12月20日，檢自：http://www.wikiwand.com/zh-hk/%E9%9B%BB%E5%AD%90%E5%85%83%E4%BB%B6
英文文獻
[1]Khizar Hayat., 2018, “Multimedia super-resolution via deep learning: A survey”, Digital Signal Processing, vol. 81, pp.198-217.
[2]Alex Krizhevsky, Ilya Sutskever, &Geoffrey E. Hinton., 2012, “ ImageNet Classification with Deep Convolutional Neural Networks”, In Neural Information Processing Systems 25.
[3]Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner., 1998, “Gradient-based learning applied to document recognition”, Proceedings of the IEEE, vol. 86(11) , pp.2278-2324.
[4]Yang Lu,Shujuan Yi,Nianyin Zeng,Yurong Liu,&YongZhang., 2017, “ Identification of rice diseases using deep convolutional neural networks”, Neurocomputing, vol. 267, pp.378-384.
[5]Xingcheng Luo,Ruihan Shen,Jian Hu,Jianhua Deng,Linji Hu, &Qing Guan., 2017, “ A Deep Convolution Neural Network Model for Vehicle Recognition and Face Recognition”, Procedia Computer Science, vol. 107, pp.715-720.
[6]Wei Liu,Dragomir Anguelov, Dumitru Erhan, Christian Szegedy,Scott Reed,Cheng-Yang Fu, &Alexander C. Berg., 2015, “ SSD: Single Shot MultiBox Detector”, ArXiv:1512.02325.
[7]P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun., 2014, “ Overfeat: Integrated recognition, localization and detection using convolutional networks”, ArXiv:1312.6229.
[8]K. Simonyan and A. Zisserman., 2015 , “ Very deep convolutional networks for large-scale image recognition”, ArXiv:1409.1556.
[9]C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich., 2015, “ Going deeper with convolutions”, IEEE Conference on Computer Vision and Pattern Recognition.
[10]C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna., 2015, “ Rethinking the Inception architecture for computer vision”, ArXiv:1512.00567.
[11]C. Szegedy, S. Ioffe, and V. Vanhoucke., 2016, “ Inception-v4, inception-resnet and the impact of residual connections on learning”, ArXiv:1602.07261.
[12]Boukaye Boubacar Traore, Bernard Kamsu-Foguem,＆Fana Tangara., 2018, “ Deep convolution neural network for image recognition”, Ecological Informatics, vol. 48, pp.257-268.
[13]Rulei Yu, &Lei Shi., 2018, “ A user-based taxonomy for deep learning visualization”, Visual Informatics, vol. 2(3), pp.147-154.
[14]Rui Zhao, Ruqiang Yan, Zhenghua Chen, Kezhi Mao, Peng Wang,＆Robert X. Gao., 2019, “ Deep learning and its applications to machine health monitoring”, Mechanical Systems and Signal Processing, vol. 115, pp.213-237.