臺灣博碩士論文加值系統

隨著科技的日新月異，工業的發展也快速進步。從最初的手工業到如今的自動化工業以及工業4.0時代，各類硬體運算能力不斷的提升，各種大數據資料庫不斷的產生，使得人工智慧技術被各個產業與學者所重視。比如醫療方面近年來COVID-19的疫情在全球肆虐，就有學者利用人工智慧模擬傳播途徑，向大眾宣導防疫知識。商業方面藉由人工智慧分析股票數據，推薦投資方向以減少虧損。生活方面使用影像辨識的人工智慧系統於各類裝置的安全鎖或停車場的車牌辨識系統。工業方面隨著智慧化工廠中所使用的機械手臂越來越泛用，因此需要具備能精準辨識零組件的人工智慧視覺系統。
卷積神經網路（Convolutional Neural Network, CNN）是目前最常見使用於影像辨識的人工智慧模型。從LeNet發展到GoogLeNet這期間，隨著卷積層數越來越深，辨識準確率也越來越高，但也導致模型的訓練與預測的時間也越來越長。另外，由於GoogLeNet採用了橫向分支結構的Inception模組，因而準確率比起純粹發展直向深度的VGG模型更高。為了解決現今相對低端的智慧型裝置沒有足夠的資料運算能力，需要提供足夠輕量的CNN模型。因此本文為了得到較短的模型訓練與預測時間，同時得到良好的辨識準確率，本文結合AlexNet直向淺層卷積層與GoogLeNet橫向分支結構的Inception模組，重新設計出TM-MA系列的九個模型，並且將電線、電容、電晶體、保險絲、電感線圈及可變電阻等六類電子元件分類資料集，應用在模型的訓練與驗證，嘗試探索增減卷積層數以及將部分卷積層改成1x1卷積(升維或降維)對模型準確率與訓練時間的影響。最後本文也比較了模型在有無預訓練兩種情況下的準確率與訓練時間。

With the rapid development of science and technology, the development of industry is also progressing rapidly. From the initial handicraft industry to today's automation industry and the era of Industry 4.0, various hardware computing capabilities have been continuously improved, and various big data databases have been continuously generated. Artificial intelligence technology is valued by various industries and scholars. For example, in the medical field, the COVID-19 epidemic has raged around the world in recent years, and some scholars have used artificial intelligence to simulate the transmission route to promote epidemic prevention knowledge to the public. In business, artificial intelligence analyzes stock data and recommends investment directions to reduce losses. In daily life, the artificial intelligence system of image recognition is used in the security lock of various devices or the license plate recognition system of the parking lot. In industry, as the robotic arms used in smart factories are becoming more and more versatile, it is necessary to have an artificial intelligence vision system that can accurately identify components.
Convolutional Neural Network (CNN) is currently the most commonly used artificial intelligence model for image recognition. During the development from LeNet to GoogLeNet, as the number of convolutional layers became deeper and deeper, the recognition accuracy became higher and higher, but it also led to the longer and longer training and prediction time of the model. In addition, since GoogLeNet uses the Inception module of the horizontal branch structure, the accuracy rate is higher than that of the VGG model that purely develops the vertical depth. In order to solve the problem that today's relatively low-end smart devices do not have enough data computing capabilities, it is necessary to provide a sufficiently lightweight CNN model. Therefore, in order to obtain a shorter model training and prediction time, and at the same time obtain a good identification accuracy, this paper combines the AlexNet vertical shallow convolutional layer and the GoogLeNet horizontal branch structure Inception module to redesign the nine TM-MA model series, and the classification data set of six types of electronic components, such as wires, capacitors, transistors, fuses, inductance coils and variable resistors, is used in training and verification models, trying to explore the effect of increasing or decreasing the number of convolutional layers and changing some of the convolutional layers to 1x1 convolutions (increasing or decreasing channel number) on model accuracy and training time. Finally, this paper also compares the accuracy and training time of the model with and without pre-training.

中文摘要....i
英文摘要.........ii
誌謝......iv
目錄...........v
表目錄.......... vi
圖目錄.......... vii
第一章 研究背景與動機............ 1
第二章 卷積神經網路技術簡介...... 4
2.1 LeNet神經網路....... 6
2.2 AlexNet神經網路....6
2.3 VGGNet神經網路.....8
2.4 GoogLeNet神經網路....9
第三章 卷積神經網路模型設計與實現....14
3.1設計理念......14
3.2模型結構......15
3.3模型理論......27
3.4實現方法.....30
第四章 模型訓練與分類預測結果......32
4.1實驗步驟.....32
第五章 結論......39
參考文獻...........40

[1] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner,” Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no.11, pp. 2278–2324, 1998.
[2]Alex Krizhevsky , Ilya Sutskever and Geoffrey E. Hinton “ImageNet Classification with Deep Convolutional Neural Networks” In Advances in Neural Information Processing Systems 25, pages 1106–1114, 2012.
[3] Karen Simonyan , Andrew Zisserman “Very Deep Convolutional Networks for Large-Scale Image Recognition” arXiv:1409.1556v6 [cs.CV] 10 Apr 2015
[4] Christian Szegedy , Wei Liu , Yangqing Jia , Pierre Sermanet , Scott Reed , Dragomir Anguelov , Dumitru Erhan , Vincent Vanhoucke and Andrew Rabinovich “Going deeper with convolutions” arXiv:1409.4842v1 [cs.CV] 17 Sep 2014
[5] Christian Szegedy , Vincent Vanhoucke , Sergey Ioffe , Jonathon Shlens and Zbigniew Wojna “Rethinking the Inception Architecture for Computer Vision” arXiv:1512.00567v3 [cs.CV] 11 Dec 2015
[6] D.O. Hebb,”The Organization of Behavior,” New York: Wiley & Sons, 1949.
[7] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov,” Dropout: A simple way to prevent neural networks from overfitting,” JMLR, vol.15, pp.1929–1958, 2014.
[8] https://www.kaggle.com/datasets/aryaminus/electronic-components
[9]K. He, X. Zhang, S. Ren, J. Sun, “Deep Residual Learning for Image Recognition,” In:CVPR, arXiv preprint arXiv:1512.03385, 2015
[10] Mingxing Tan , Quoc V. Le “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks” arXiv:1905.11946v5 [cs.LG] 11 Sep 2020
[11] Kumar Chellapilla ,Sidd Puri and Patrice Simard“High Performance Convolutional Neural Networks for Document Processing” ,2006.
[12]Yao-Yuan Xu “Convolutional Neural Network Models for Flowers Recognition”,2021.
[13]https://www.image-net.org/download.php