從 2019 年 12 月起因新型冠狀病毒所造成的疾病 COVID-19 (Coronavirus disease 2019)爆發以來，截止到目前 2022 年 6 月，台灣已累計確診 319 萬個病例數。雖然當前台灣的人口疫苗覆蓋率第一劑達九成，第兩劑有八成二，第三劑也有六成九，有相對的基礎保護力。但為了避免本土疫情持續升溫，目前政府的防疫政策除例外情形時免戴口罩，外出仍應全程佩戴口罩，相關規定持續至 2022 年 6 月底。因應疫情下的防疫措施，本論文設計了一套人臉口罩辨識模型，可用來辨別是否有佩戴口罩以及是否有佩戴好口罩。
本論文運用了深度學習網路，採用了物件偵測模型為 YOLO(You Only Look Once)v4，相對於歷代的 YOLO 版本不僅提升了檢測速度，且精準度也大幅改善，再搭配生成對抗網路模型來做資料擴增，豐富了訓練的資料集項目，來提升辨識的準確度。採用生成對抗網路模型的資料集對比沒有經過處理的資料集平均預測率可提升 5%，達到 94.52%，召回率為 93%，F1-Score 為 91%。

The COVID-19 (Coronavirus disease 2019) pandemic broke out in December 2019. Until now, in June 2022, there are in total of 3.19 million confirmed cases in Taiwan. The current rate of vaccination of the Taiwanese population seems adequate to provide basic immune protection against the virus. 90% of Taiwanese people had their first COVID-19 vaccine, 82% of which got the second vaccination and 69% of the population completed their booster dose. However, to prevent the pandemic from getting worse, except few special situations, all individuals should wear masks while going outdoor due to government policies, several related regulations will remain until the end of June. To cope with the mask regulations, this study designed a facial mask recognition model, the model can be used to examine whether an individual wears a mask and if a mask is worn properly by one.
This design applied Deep Learning Network and selected YOLO (You Only Look Once) v4 as the object detection model. This version of YOLO, compared to the past few versions of it, not only speeds up the detection but also significantly enhanced its accuracy. Working with the Generative Adversarial Network for Data Augmentation, the database for detection was enriched and hence reached a greater level of accuracy. The average prediction rate of data collections can go up to 94.52%, 5%more than the rate not applied with Generative Adversarial Network, with the recall rate of 93% and F1-score of 91%.

[1] 疾病介紹, (2020, April 27), 衛生福利部疾病管制署
https://www.cdc.gov.tw/En/Category/Page/vleOMKqwuEbIMgqaTeXG8A
(Access Date: August 30, 2022)
[2] Covid-19, (2016, August 21), 衛生福利部疾病管制署. https://sites.google.com/cdc.gov.tw/2019ncov/global
(Access Date: August 30, 2022)
[3] Larxel, (2020, Summer 5), Face Mask Detection. Kaggle, https://www.kaggle.com/datasets/andrewmvd/face-mask-detection.
(Access Date: August 30, 2022)
[4] I. Goodfellow, J Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville and Y. Bemgio, "Generative Adversarial Networks", 2014, doi:
https://doi.org/10.48550/arXiv.1406.2661.
[5] M. Mirza, S.Osindero, "Conditional Generative Adversarial Nets", 2014,
doi: https://doi.org/10.48550/arXiv.1411.1784.
[6] A. Radfod, L. Metz, S.Chintala, "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks", doi:
https://doi.org/10.48550/arXiv.1511.06434.
[7] T. Shaham, T. Dekel, and T. Michaeli, " SinGAN: Learning a Generative Model from a Single Natural Image", 2019, pp 4570-4572,
https://openaccess.thecvf.com/content_ICCV_2019/papers/Shaham_SinGAN_Learning_a_Generative_Model_From_a_Single_Natural_Image_ICCV_2019_paper.pdf
[8] T. Hinz, M. Fisher, O. Wang, S. Wermter, "Improved Techniques for Training Single-Image GANs", 2020 , pp. 2-3, doi:
https://doi.org/10.48550/arXiv.2003.11512
[9] M. Childs, (2011, November 1), John McCarthy: Computer scientist known as the father of AI.
https://www.independent.co.uk/news/obituaries/john-mccarthy-computer-scientist-known-as-the-father-of-ai-6255307.html
(Access Date: August 30, 2022)
[10] Y. Liu. (2020, July 19). AI 學習筆記 (I) 機器學習 vs. 深度學習. Medium.
https://medium.com/mr-lius-murmur/ai-%E5%AD%B8%E7%BF%92%E7%AD%86%E8%A8%98-i-%E6%A9%9F%E5%99%A8%E5%AD%B8%E7%BF%92-vs-%E6%B7%B1%E5%BA%A6%E5%AD%B8%E7%BF%92-98b1ce9123a3
(Access Date: August 30, 2022)
[11] R. H. Wurtz, (2013, November 1), David H. Hubel (1926–2013), Science.
https://www.science.org/doi/10.1126/science.1247113.
(Access Date: August 30, 2022)
[12] K. Fukushima, (1980), " Neocognitron : A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position". Biological Cybernetics. 36 (4): 193–202. doi:10.1007/bf00344251. PMID 7370364. S2CID 206775608.
[13] Y. Lecun, L. Bottou, Y. Bengio and P. Haffner, "Gradient-based learning applied to document recognition,” in Proceedings of the IEEE, vol. 86, no. 11, pp. 6-7, Nov 1998, doi: 10.1109/5.726791
[14] R. Girshick, J. Donahue, T. Darrell, & J. Malik, "Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation", 2014, pp.1, doi:
https://doi.org/10.48550/arXiv.1311.2524
[15] R. Girshick. "Fast R-CNN", 2015, pp.2, doi:
https://doi.org/10.48550/arXiv.1504.08083
[16] J. Redmon (2015) Joseph Redmon website
https://pjreddie.com/
(Access Date: August 30, 2022)
[17] J. Redmon, S. Divvala, R. Girshick, & A. Farhadi, "You Only Look Once: Unified, Real-Time Object Detection", 2016, pp.1–3, doi:
https://doi.org/10.48550/arXiv.1506.02640
[18] J. Redmon, & A. Farhadi, "YOLO9000: Better, Faster, Stronger.", 2016, pp. 4, doi:
https://doi.org/10.48550/arXiv.1506.02640
[19] J. Redmon, & A. Farhadi, "YOLOv3:AnIncremental Improvement", 2018, pp.2, doi:
https://doi.org/10.48550/arXiv.1804.02767
[20] A. Bochkovskiy, C.Y. Wang, & H. Liao, "YOLOv4: Optimal Speed and Accuracy of Object Detection", 2020, pp.1-2, doi:
https://doi.org/10.48550/arXiv.2004.10934
[21] 江伶娸, "結合CGAN與YOLOv4進行芒果等級分類", 朝陽科技大學資訊管理系碩士論文, 2022, pp.22, doi:
https://hdl.handle.net/11296/phqyup
[22] P. Wu, (2020, September 7), Object Detection: 初探YOLOV5爭議以及與YOLOV4之比較, Medium,
https://pedin024.medium.com/%E5%88%9D%E6%8E%A2yolov5-71f13b4ba78d
(Access Date: August 30, 2022)
[23] K. Zhang, Z. Zhang, & Z. Li., "Joint Face Detection and Alignment Using Multi-Task Cascaded Convolutional Networks", 2016, pp.1–3, doi:
https://doi.org/10.1109/LSP.2016.2603342
[24] J. Solawetz, (2020, May 13), Roboflow
https://blog.roboflow.com/yolov4-data-augmentation/
(Access Date: August 30, 2022)
[25] 李謦伊, (2020, August 6), YOLOv4 訓練教學, Medium.
https://medium.com/ching-i/yolo-c49f70241aa7
(Access Date: August 30, 2022)