現代的各項產品都存在著瑕疵風險，為避免這類的風險存在，在產品出廠前都會進行嚴密的瑕疵檢測，以確保產品的品質。傳統工廠中，大多都是採用人工目檢的方式來檢查產品是否有瑕疵，但檢測過程中不僅耗時、耗力，人員也會因為疲勞而有LOSS的風險，為避免這類情況發生，進而衍生出自動化影像辨識系統，針對產品做出各項檢測，檢測該產品是否有瑕疵風險。
在本論文中，我利用JETSON NANO製作一個玻璃瑕疵檢測模型，藉由玻璃各類影像圖進行預測辨識之途，由於影像圖是我自行拍攝，在訓練途中，容易有辨識率偏低的狀況。因此本論文在第一部分會先進行圖像處理，讓整理過的影像圖能夠更好的訓練模型；再來第二部分是使用目前最新的YOLOV5演算法，以性能上來說YOLOV5是略顯弱於YOLOV4，但是靈活性與速度上是高於YOLOV4，於是我便選擇使用YOLOV5來做為我影像辨識的演算模型。
玻璃為易碎品，只要有些微的表面缺陷，就容易有破裂的風險，進而導致購買的顧客有受傷的可能性，本研究以玻璃為檢測目標，在拍攝玻璃缺陷時，容易有反射光以及光源入射角的問題，導致拍攝時玻璃缺陷會不夠明顯，透過不斷嘗試，發現越近距離拍攝，反光的面積會減少越多，在以傾斜的角度拍攝，玻璃瑕疵顯現度會較高，讓機器在訓練時，效果會更好。

All modern products have the risk of defects. In order to avoid such risks, strict defect inspections are carried out before the products leave the factory to ensure the quality of the products. In traditional factories, manual visual inspection is mostly used to check whether products are defective. However, the inspection process is not only time-consuming and labor-intensive, but also personnel are at risk of LOSS due to fatigue. An automated image recognition system is derived to perform various tests on the product to detect whether the product is at risk of defects.
In this paper, I use JETSON NANO to make a glass defect detection model, and use various image images of glass for prediction and identification. Since the image images are taken by myself, it is easy to have a low recognition rate during training. Therefore, in the first part of this thesis, image processing will be performed first, so that the sorted images can better train the model; then the second part is to use the latest YOLOV5 algorithm. In terms of performance, YOLOV5 is slightly weaker than YOLOV4. , but the flexibility and speed are higher than YOLOV4, so I chose to use YOLOV5 as the algorithm model for my image recognition.
Glass is a fragile product. As long as there are some slight surface defects, it is easy to have the risk of rupture, which will lead to the possibility of injury to customers who buy it. This research takes glass as the detection target. When shooting glass defects, it is easy to have reflected light and The problem of the incident angle of the light source makes the glass defects not obvious enough when shooting. Through continuous attempts, it is found that the closer the shooting is, the more the reflective area will be reduced. When shooting at an oblique angle, the visibility of glass defects will be higher. When training, the effect will be better.

目錄
中文摘要 i
英文摘要 ii
誌 謝 iii
目錄 iv
圖目錄 vi
表目錄 viii
第1章 緒論 - 1 -
1.1 研究背景 - 1 -
1.2 研究目的 - 3 -
1.3 論文架構 - 3 -
第2章文獻回顧 - 4 -
2.1機器學習 - 4 -
2.1.1監督式學習 - 4 -
2.1.2非監督式學習 - 4 -
2.1.3半監督式學習 - 4 -
2.1.4強化學習 - 5 -
2.2.深度學習 - 5 -
2.2.1卷積神經網路(CNN) - 6 -
2.2.2循環神經網路(RNN) - 7 -
2.3 YOLOV系列 - 8 -
2.3.1 YOLOV1 - 8 -
2.3.2 YOLOV2 - 10 -
2.3.3 YOLOV3 - 13 -
2.3.4 YOLOV4 - 15 -
2.3.5 YOLOV5 - 22 -
第3章 系統整合 - 24 -
3.1硬體介紹 - 24 -
3.1.1 Jetson Nano - 24 -
3.1.2 GPIO接腳 - 27 -
3.2軟體介紹 - 29 -
3.2.1 Jetson Nano作業系統 - 29 -
3.2.2安裝程式 - 31 -
3.2.3 Labelimg - 31 -
第4章實驗結果 - 35 -
4.1實驗流程 - 35 -
4.2實驗結果分析 - 37 -
第5章 結論與未來展望 - 46 -
5.1結論 - 46 -
5.2未來展望 - 46 -
參考文獻 - 47 -

1.McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (2006). A proposal for the dartmouth summer research project on artificial intelligence, august 31, 1955. AI magazine, 27(4), 12-12.
2.Michalski, R. S., Carbonell, J. G., & Mitchell, T. M. (Eds.). (2013). Machine learning: An artificial intelligence approach. Springer Science & Business Media.
3.LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. nature, 521(7553), 436-444.
4.偉育TibaMeBLOG Q13人工智慧(AI)、機器學習和深度學習有甚麼區別?Available: https://blog.tibame.com/?p=17580
5.JetsonHacks Developing for NVIDIA Jetson Available:https://jetsonhacks.com/nvidia-jetson-nano-j41-header-pinout/
6.Cunningham, P., Cord, M., & Delany, S. J. (2008). Supervised learning. In Machine learning techniques for multimedia (pp. 21-49). Springer, Berlin, Heidelberg.
7.Ghahramani, Z. (2003, February). Unsupervised learning. In Summer school on machine learning (pp. 72-112). Springer, Berlin, Heidelberg.
8.Zhu, X. J. (2005). Semi-supervised learning literature survey.
9.Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction. MIT press.
10.偉育TibaMeBLOG Q9什麼是機器學習? Available:https://blog.tibame.com/?p=18569
11.O'Shea, K., & Nash, R. (2015). An introduction to convolutional neural networks. arXiv preprint arXiv:1511.08458.
12.LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324.
13.Analytics Vidhya The Architecture of Lenet-5 Available: https://www.analyticsvidhya.com/blog/2021/03/the-architecture-of-lenet-5/
14.Zaremba, W., Sutskever, I., & Vinyals, O. (2014). Recurrent neural network regularization. arXiv preprint arXiv:1409.2329.
15.Balachandran, S. (2009). General Purpose Input/Output (GPIO). Michigan State University College of Engineering. Published, 08-11.
16.Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), 1735-1780.
17.Ashing’s Blog 深度學習(3)--循環神經網絡(RNN,Recurrent Neural Networks) Available:https://arbu00.blogspot.com/2017/05/3-rnn-recurrent-neural-networks.html
18.Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 779-788).
19.Hosang, J., Benenson, R., & Schiele, B. (2017). Learning non-maximum suppression. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4507-4515).
20.Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 779-788).
21.Redmon, J., & Farhadi, A. (2017). YOLO9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 7263-7271).
22.Ioffe, S., & Szegedy, C. (2015, June). Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning (pp. 448-456). PMLR.
23.Tian, Z., Shen, C., Chen, H., & He, T. (2019). Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 9627-9636).
24.NVIDIA DEVELOPER Getting Stared with Jetson Nano Developer Kit
Available:https://developer.nvidia.com/embedded/learn/get-started-jetson-nano-devkit#intro
25.Redmon, J., & Farhadi, A. (2018). Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767.
26.Lin, T. Y., Dollár, P., Girshick, R., He, K., Hariharan, B., & Belongie, S. (2017). Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2117-2125).
27.Kleinbaum, D. G., Dietz, K., Gail, M., Klein, M., & Klein, M. (2002). Logistic regression (p. 536). New York: Springer-Verlag.
28.Bochkovskiy, A., Wang, C. Y., & Liao, H. Y. M. (2020). Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934.
29.Wang, C. Y., Liao, H. Y. M., Wu, Y. H., Chen, P. Y., Hsieh, J. W., & Yeh, I. H. (2020). CSPNet: A new backbone that can enhance learning capability of CNN. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (pp. 390-391).
30.Xie, S., Girshick, R., Dollár, P., Tu, Z., & He, K. (2017). Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1492-1500).
31.He, K., Zhang, X., Ren, S., & Sun, J. (2015). Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9), 1904-1916.
32.Liu, S., Qi, L., Qin, H., Shi, J., & Jia, J. (2018). Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 8759-8768).
33.Yun, S., Han, D., Oh, S. J., Chun, S., Choe, J., & Yoo, Y. (2019). Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 6023-6032).
34.Santhakumar, K., Kiran, B. R., Gauthier, T., & Yogamani, S. (2021). Exploring 2D data augmentation for 3D monocular object detection. arXiv preprint arXiv:2104.10786.
35.Ghiasi, G., Lin, T. Y., & Le, Q. V. (2018). Dropblock: A regularization method for convolutional networks. Advances in neural information processing systems, 31.
36.Müller, R., Kornblith, S., & Hinton, G. E. (2019). When does label smoothing help?. Advances in neural information processing systems, 32.
37.Misra, D. (2019). Mish: A self regularized non-monotonic activation function. arXiv preprint arXiv:1908.08681.
38.Tan, M., & Le, Q. (2019, May). Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning (pp. 6105-6114). PMLR.
39.Woo, S., Park, J., Lee, J. Y., & Kweon, I. S. (2018). Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV) (pp. 3-19).
40.NVIDIA DEVELOPER Jetson Nano Developer Kit Available:https://developer.nvidia.com/embedded/jetson-nano-developer-kit
41.Ultralytics/yolov5—v6.1 – TensorRT,Tensorflow Edge TPU and OpenVINO Export and Inference。Available: https://github.com/ultralytics/yolov5/releases
42.楊安琪(2017)。審查文件只需幾秒!摩根大通AI軟體可替代律師年省36萬小時。Available:https://technews.tw/2017/03/03/jpmorgan-software-coin-reviews-documents-in-seconds/(Jun.20,2022).