中醫辨證常以望、聞、問、切四診做為依據，而舌診是望診中主要的內容，由仔細觀察舌頭的表象即可找出病患的體質與可能的病因。本研究利用人工智慧的技術來處理影像辨識，建立了一套自動且有效率建立了一套自動且有效率的即時舌診影像辨別系統。且針對舌象的取樣與分析，找出舌特徵之間的特性，經過數百筆診斷資料的訓練，並請專業的醫師進行確認後，本系統對於齒痕舌與裂紋舌的舌象的辨識正確率可高達平均九成以上。若在未來的研究中能夠加入其他中醫診斷系統的分析結果，將可提供中醫師們更詳盡的病情訊息，進而輔助中醫師們做出更精準的病症診斷。本論文兩部分的研究中，資料集內兩種標籤的數量雖不平衡，但實驗結果表明仍有一定的表現，說明不平衡的舌頭資料集還是可以在不特別處理的情況下透過深度學習有一些作用。
針對舌診數位影像的辨識，本研究經過比較多種深度學習的演算法，最終採用YOLO v4神經網路訓練模型，建立一套可快速進行物件判別的舌診圖像辨識系統。經過實驗結果，我們確定不同解析度的舌部影像會對模型的判斷產生非常大幅的影響；當舌部影像的解析度小於128x96，模型判別的表現將開始下滑，隨著解析度的下降模型的判別能力也隨之降低，最後模型將喪失任何的判斷能力。因此須確保所有來源之舌部影像的解析度大小，方可正確地辨識舌象圖形，提供中醫師更詳盡的病情訊息，由模型自動判斷舌部的症狀來減少人力的消耗，進而輔助中醫師做出更精準的病症診斷。

Chinese medicine is often based on the four diagnoses of looking, smelling, asking and cutting, and tongue diagnosis is the main component of looking. The physical condition of the patient and the possible cause of the disease can be identified by careful observation of the appearance of the tongue. This study uses artificial intelligence to process image recognition, and establishes an automatic and efficient real-time tongue diagnosis image recognition system. After training hundreds of diagnostic data and asking professional doctors to confirm, the system can achieve an average accuracy rate of more than 90% for some tongue symptoms. If the analysis results of other TCM diagnostic systems can be included in future studies, it will provide TCM practitioners with more detailed information about their conditions, which will in turn assist them in making more accurate diagnoses.
After comparing various deep learning algorithms for tongue diagnosis, this study uses the YOLO v4 neural network training model to build a tongue recognition system for rapid object discrimination. As a result of our experiments, we determined that different resolutions of tongue images can have a very significant effect on the judgment of the model; When the resolution of the tongue image is less than 128x96, the model's discrimination performance will start to decline, and as the resolution decreases, the model's discrimination ability will also decrease, and eventually the model will lose any judgment ability. Therefore, it is necessary to ensure the resolution of tongue images from all sources in order to correctly identify tongue images and provide TCM practitioners with more detailed information about their conditions.

第一章 緒論 1
1.1 研究背景 1
1.2 研究動機與目的 2
1.3 論文架構 3
第二章 相關研究與文獻探討 4
2.1 相關研究介紹 4
2.1.1 舌部影像與物件偵測 4
2.1.2 自動化舌診系統 4
2.2 常見中醫舌象理論介紹 5
2.3 卷積神經網路 6
2.4 物件偵測深度學習框架YOLOv4 10
2.4.1 YOLOv4簡介 10
2.4.2 YOLOv4基本架構 12
2.5 評估方式 14
2.5.1 IoU （Intersection over Union） 14
2.5.2 混淆矩陣（Confusion Matrix） 16
2.5.3 mAP（mean Average Precision） 19
第三章 研究方法 20
3.1 研究設備與環境 20
3.2 資料集 21
3.2.1 普通舌的資料集 22
3.2.2 齒痕舌的資料集 23
3.2.3 裂痕舌的資料集 24
3.3 舌象辨識－普通舌，裂痕舌，齒痕舌 25
3.3.1 訓練資料預處理 27
3.3.2 模型訓練 28
第四章 實驗設計與結果觀察 30
4.1 普通舌、齒痕舌、裂痕舌之模型訓練過程與辨識結果 30
4.1.1 使用yolov4權重的模型訓練過程 31
4.1.2 yolov4模型辨識結果觀察 33
4.1.3 使用yolov4-tiny權重的模型訓練和辨識結果 41
4.2 不同尺寸之圖像對於模型辨識之影響 67
4.2.1 資料預處理 69
4.2.2 各個像素尺寸級距的ROC Curve圖表 71
4.2.3 全像素級距的Average AUC的觀察比較 77
第五章 結論與未來展望 78

[1]Bochkovskiy, A., Wang, C.-Y., & Liao, H.-Y. M. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. ArXiv:2004.10934 [Cs, Eess]. Retrieved from https://arxiv.org/abs/2004.10934
[2]AlexeyAB (2020) darknet [Source code]. https://github.com/AlexeyAB/darknet.
[3]Joseph Redmon. (2013--2016). Darknet: Open Source Neural Networks in C. Retrieved from https://pjreddie.com/darknet/ (Nov. 4, 2021)
[4]中國舌診大全. 李乃民上海. 学苑出版社page 1610 ss:10226112 1994/02
[5]舌診研究. v2. 陳澤霖. 陳梅芳. 上海科學技術出版社1982/12
[6]舌診現代化研究的回顧與展望. 劉慶. 岳小強. 淩昌全. Journal of Chinese Integrative Medicine, May 2003 , Vol .1 , No .1
[7]Wang, X., Liu, J., Wu, C., Liu, J., Li, Q., Chen, Y., … Chen, J. (2020). Artificial intelligence in tongue diagnosis: Using deep convolutional neural network for recognizing unhealthy tongue with tooth-mark. Computational and Structural Biotechnology Journal, 18, 973–980. https://doi.org/10.1016/j.csbj.2020.04.002
[8]Hsu, Y.-C., Chen, Y.-C., Lo, L., & Chiang, J. Y. (2010, December 1). Automatic tongue feature extraction. https://doi.org/10.1109/COMPSYM.2010.5685377
[9]王彥暉. (2015). 臨床實用舌象圖譜. 新北市：合記圖書出版社
[10]YOLO: Real-Time Object Detection.
https://pjreddie.com/darknet/yolo/
[11]YOLOv4 model architecture
https://iq.opengenus.org/yolov4-model-architecture/
[12]國家高速網路與計算中心. NARLabs財團法人國家實驗研究院
https://www.nchc.org.tw/
[13]tsutalin. LabelImg
https://github.com/tzutalin/labelImg
[14]Lo, L.-C., Hou, C.-C., Chen, Y.-L., Chiang, J., & Hsu, C.-C. (n.d.). Automatic Tongue Diagnosis System. Retrieved from https://image.cse.nsysu.edu.tw/ServerPaper/52/Automatic%20Tongue%20Diagnosis%20System.pdf
[15]Hu, M.-C., Lan, K.-C., Fang, W.-C., Huang, Y.-C., Ho, T.-J., Lin, C.-P., … Tseng, K.-C. (2017). Automated tongue diagnosis on the smartphone and its applications. Computer Methods and Programs in Biomedicine. https://doi.org/10.1016/j.cmpb.2017.12.029
[16]Alzubaidi, L., Zhang, J., Humaidi, A. J., Al-Dujaili, A., Duan, Y., Al-Shamma, O., … Farhan, L. (2021). Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. Journal of Big Data, 8(1). https://doi.org/10.1186/s40537-021-00444-8
[17]Albawi, S., Mohammed, T. A., & Al-Zawi, S. (2017). Understanding of a convolutional neural network. 2017 International Conference on Engineering and Technology (ICET). https://doi.org/10.1109/icengtechnol.2017.8308186
[18]Tan, M., Pang, R., & Le, Q. V. (2020). EfficientDet: Scalable and Efficient Object Detection. ArXiv:1911.09070 [Cs, Eess]. Retrieved from https://arxiv.org/abs/1911.09070
[19]James, Y. (2021, June 26). [資料分析&機器學習] 第5.1講: 卷積神經網絡介紹(Convolutional Neural Network). Retrieved from JamesLearningNote website: https://miro.medium.com/max/700/1*irWQaiIjHS27ZAPaVDoj6w.png
[20]Convolutional Neural Networks (CNN)- Step 1- Convolution Operation. (2018, December 28). Retrieved August 19, 2022, from Mohamed DHAOUI website: https://sds-platform-private.s3-us-east-2.amazonaws.com/uploads/70_blog_image_3.png
[21]Convolutional Neural Networks (CNN)- Step 1- Convolution Operation. (2018, December 28). Retrieved August 19, 2022, from Mohamed DHAOUI website: https://sds-platform-private.s3-us-east-2.amazonaws.com/uploads/70_blog_image_4.png
[22]Convolutional Neural Networks (CNN)- Step 1- Convolution Operation. (2018, December 28). Retrieved August 19, 2022, from Mohamed DHAOUI website: https://sds-platform-private.s3-us-east-2.amazonaws.com/uploads/70_blog_image_5.png
[23]Tamse, V. U. (2020, February 20). Understanding Convolutional Neural Networks (CNN or ConvNet). Retrieved August 19, 2022, from Analytics Vidhya website: https://miro.medium.com/max/700/1*FPCP5wlOuRjdUoODpdsEmw.png
[24]卷積神經網路 Convolutional Neural Networks · 資料科學・機器・人. (n.d.). Retrieved from brohrer.mcknote.com website: https://youtu.be/FmpDIaiMIeA?t=14m03s
[25]Bochkovskiy, A., Wang, C.-Y., & Liao, H.-Y. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. Figure 1. [Image]. Retrieved from https://arxiv.org/pdf/2004.10934.pdf
[26]Bochkovskiy, A., Wang, C.-Y., & Liao, H.-Y. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. Figure 2. Object detector. [Image]. Retrieved from https://arxiv.org/pdf/2004.10934.pdf
[27]Bochkovskiy, A., Wang, C.-Y., & Liao, H.-Y. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. Table 1. Parameters of neural networks for image classification. [Table]. Retrieved from https://arxiv.org/pdf/2004.10934.pdf
[28]‌Intersection over Union (IoU) for object detection. (2016, November 7). Retrieved from PyImageSearch website: https://929687.smushcdn.com/2633864/wp-content/uploads/2016/09/iou_stop_sign.jpg?lossy=1&strip=1&webp=1
[29]Intersection over Union (IoU) for object detection. (2016, November 7). Retrieved from PyImageSearch website: https://929687.smushcdn.com/2633864/wp-content/uploads/2016/09/iou_equation.png?lossy=1&strip=1&webp=1
[30]Cowton, J., Kyriazakis, I., & Bacardit, J. (2019). Automated Individual Pig Localisation, Tracking and Behaviour Metric Extraction Using Deep Learning. IEEE Access, 7, 108049–108060. Figure 6. Examples of how Intersection over Union is calculated. Left: Poor performance IoU = 0.4034. Middle: Good performance, IoU = 0.7330. Right: Excellent performance, IoU = 0.9264. [Image] https://doi.org/10.1109/access.2019.2933060
[31]Kayyath, A. (2021, October 28). Confusion Matrix : Let’s clear this confusion. Retrieved August 19, 2022, from Medium website: https://miro.medium.com/max/802/1*Kn0-69rgSKIQzdFN1czpZw.png
[32]MaDi. (2020, December 4). 機器學習-常見的評估指標. Retrieved August 19, 2022, from MaDi’s Blog website: https://i.imgur.com/EoyROPJ.png
[33]iThome. (n.d.). iT 邦幫忙::一起幫忙解決難題，拯救 IT 人的一天. Retrieved August 19, 2022, from iT 邦幫忙::一起幫忙解決難題，拯救 IT 人的一天 website: https://ithelp.ithome.com.tw/upload/images/20191117/20001976fKvLEvonQR.png