台灣是從製造業起家的國家；在各類技術如此進步的21世紀中無論工廠流水線如何演化，其產品良率一直都是必須解決的重點問題。將不良品檢測出來並銷毀需要一定的人力及成本，如果將不良品送至廠商或是送至消費者手上無疑會造成更大的損失，最嚴重更可能導致人員傷亡。傳統製造業中，會雇用大量的人力來確保產品的品質，但由於員工精力無法時刻皆保持穩定因此可能導致不良的產品流出，這無疑造成重大的損失。現今的高科技產業主要使用AOI(Automated Optical Inspection)來針對產品進行瑕疵檢測，而所謂的AOI是透過光學儀器以高速度且高精確度的模式，利用人工智慧機械視覺技術來達成檢測目標。不過除了AOI以外，本研究亦嘗試使用YOLO技術做為產品檢測的方法。技術基礎上，YOLO是透過卷積神經網路來進行學習的，將大量的圖像資料經過卷積以後由全連接層進行輸出來辨別產品的缺陷。技術比較上，AOI依賴於使用人為設定好的參數去進行辨識，而YOLO則是利用學習獲得的模型進行辨識。如此可推斷AOI比較適合短期的生產項目，而YOLO則可以利用於長期生產的產品。因為雖然YOLO需要大量的資料來進行訓練獲得模型，不過訓練好的YOLO模型便可以透過不斷的學習來提升分辨率，這就是只能依賴人工調整參數的AOI技術望塵莫及之處了。總結來說現今的製造業動輒生產數量龐大的訂單，在使用深度學習的YOLO時會更容易展現出其長處，而AOI則比較適合用於檢測數量較少的訂單，並需要人工調整參數，短期內AOI的分辨率會比較高。而本研究之目的就是利用YOLO技術的特點來針對醫療產品硬體部分進行主要的瑕疵檢測。

The manufacturing industry is the economic foundation of Taiwan. The development of every technology is so impressive in the 21st century as we should see. Even though the performance of product lines is getting improved, the yield rate of products is still the main issue. The mission of defect detection among products would cost significant labor sources and work time. The products with defects would cause the manufacturers or consumers further losses. Those products would probably result in injuries or even deaths in some situations. In traditional manufacturing, employers would like to employ a large amount of labor intervention to ensure the quality of products. However, because employees' energy is not always stable, it may cause the outflow of defective products, resulting in significant losses. High-tech industries currently use AOI (Automated Optical Inspection) to detect defects in products. AOI is achieved by high-speed and high-precision optical instruments, along with artificial intelligence of machine vision technology. In this research, we used the YOLO technique to develop a product detection method. The YOLO framework uses convolutional neural networks to learn how to identify defective products. The convolutional kernels deal with input images, and then the fully connected layer outputs the predicted out classified results. The AOI technique technically relies on the pre-set parameters for identification, whereas YOLO uses a well-trained model obtained by learning from data. The YOLO technique needs a large number of data to train its model, and then the well-trained model can continuously improve the ability of identification through sustained learnings. That is the reason that participants believe the YOLO technique outperforms the AOI. In conclusion, it is easier for YOLO to show its strengths in massive production lines, while AOI is suitable for smaller production lines. The purpose of this research thesis is to develop a defect detection method by utilizing the YOLO technique. Hopefully, we can extend the developed model to applications on medical products.

[1] Solomonoff, R.J.The Time Scale of Artificial Intelligence; Reflections on Social Effects, Human Systems Management, Vol 5 1985, Pp 149-153.
[2] O'Regan, G. (2018). EDVAC and ENIAC Computers.
[3] Newell, A.; Shaw, J.C.; Simon, H.A. (1959). Report on a general problem-solving program. Proceedings of the International Conference on Information Processing. pp. 256–264.
[4] Shapiro, Ehud Y. "The fifth generation project—a trip report." Communications of the ACM 26.9 (1983): 637-641.
[5] Goldberg, David E.; Holland, John H. (1988). "Genetic algorithms and machine learning" (PDF). Machine Learning. 3 (2): 95–99. doi:10.1007/bf00113892.
[6] Lewis, Michael B; Ellis, Hadyn D (2003), "How we detect a face: A survey of psychological evidence", International Journal of Imaging Systems and Technology, 13: 3–7, doi:10.1002/ima.10040, S2CID 14976176.
[7] M. Dikmen and C. Burns, "Trust in autonomous vehicles: The case of Tesla Autopilot and Summon," 2017 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2017, pp. 1093-1098, doi: 10.1109/SMC.2017.8122757.
[8] Rossion, Bruno; Hanseeuw, Bernard; Dricot, Laurence. Defining face perception areas in the human brain: A large-scale factorial fMRI face localizer analysis. Brain and Cognition. 2012-7, 79 (2): 138–157 [2019-09-13]. doi:10.1016/j.bandc.2012.01.001.
[9] Krizhevsky, A., Sutskever, I., & Hinton, G.E. (2012). ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60, 84 - 90.
[10] Bergstra, J., Bastien, F., Breuleux, O., Lamblin, P., Pascanu, R., Delalleau, O., Desjardins, G., Warde-Farley, D., Goodfellow, I., Bergeron, A., & Bengio, Y. (2012). Theano: Deep Learning on GPUs with Python.
[11] Robbins, Martin. Does an AI need to make love to Rembrandt's girlfriend to make art?. The Guardian. 2016-05-06 [2016-06-22].
[12] Redmon, Joseph et al. “You Only Look Once: Unified, Real-Time Object Detection.” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016): 779-788.
[13] Krizhevsky, A., Ilya Sutskever and Geoffrey E. Hinton. “ImageNet classification with deep convolutional neural networks.” Communications of the ACM 60 (2012): 84 - 90.
[14] Redmon, Joseph and Ali Farhadi. “YOLO9000: Better, Faster, Stronger.” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017): 6517-6525.
[15] Redmon, Joseph and Ali Farhadi. “YOLOv3: An Incremental Improvement.” ArXiv abs/1804.02767 (2018): n. pag.
[16] Lin, Tsung-Yi, Piotr Dollár, Ross B. Girshick, Kaiming He, Bharath Hariharan and Serge J. Belongie. “Feature Pyramid Networks for Object Detection.” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017): 936-944.
[17] Bochkovskiy, Alexey, Chien-Yao Wang and H. Liao. “YOLOv4: Optimal Speed and Accuracy of Object Detection.” ArXiv abs/2004.10934 (2020): n. pag.
[18] Zheng, Zhaohui, P. Wang, Wei Liu, Jinze Li, Rongguang Ye and Dongwei Ren. “Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression.” ArXiv abs/1911.08287 (2020): n. pag.
[19] Schmidhuber, J.. “Deep learning in neural networks: An overview.” Neural networks : the official journal of the International Neural Network Society 61 (2015): 85-117.
[20] https://aidea-web.tw/topic/285ef3be-44eb-43dd-85cc-f0388bf85ea4(2021).
[21] Wang, Chien-Yao, Alexey Bochkovskiy and H. Liao. “Scaled-YOLOv4: Scaling Cross Stage Partial Network.” ArXiv abs/2011.08036 (2020): n. pag.
[22] He, Kaiming, X. Zhang, Shaoqing Ren and Jian Sun. “Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition.” IEEE Transactions on Pattern Analysis and Machine Intelligence 37 (2015): 1904-1916.
[23] Misra, Diganta. “Mish: A Self Regularized Non-Monotonic Activation Function.” BMVC (2020).
[24] Orchard, Jeff and Craig S. Kaplan. “Cut-out image mosaics.” NPAR (2008).
[25] Singh, Krishna Kumar, Hao Yu, Aron Sarmasi, G. Pradeep and Yong Jae Lee. “Hide-and-Seek: A Data Augmentation Technique for Weakly-Supervised Localization and Beyond.” ArXiv abs/1811.02545 (2018): n. pag.
[26] Chen, Pengguang, Shu Liu, Hengshuang Zhao and Jiaya Jia. “GridMask Data Augmentation.” ArXiv abs/2001.04086 (2020): n. pag.
[27] Yun, Sangdoo, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe and Y. Yoo. “CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features.” 2019 IEEE/CVF International Conference on Computer Vision (ICCV) (2019): 6022-6031.
[28] Dubey, A. and Vanita Jain. “Comparative Study of Convolution Neural Network’s Relu and Leaky-Relu Activation Functions.” (2019).
[29] Ghiasi, Golnaz, Tsung-Yi Lin and Quoc V. Le. “DropBlock: A regularization method for convolutional networks.” NeurIPS (2018).
[30] Liu, Shu, Lu Qi, Haifang Qin, J. Shi and J. Jia. “Path Aggregation Network for Instance Segmentation.” 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (2018): 8759-8768.
[31] Rezatofighi, S. H., Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, I. Reid and S. Savarese. “Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression.” 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019): 658-666.
[32] http://faculty.neu.edu.cn/yunhyan/NEU_surface_defect_database.html(2021).
[33] Redmon, J., & Farhadi, A. (2018). YOLOv3: An Incremental Improvement. ArXiv, abs/1804.02767.