臺灣博碩士論文加值系統

在現代社會中，咖啡作為歷史悠久的飲料，已經成為多數人日常生活中不可或缺的飲品，如何確保其穩定的口感是非常重要的。常見的挑選咖啡生豆方式為人工挑選與色選機挑選，人工挑選是一件耗費時間也耗費精神的工作；而色選機挑選速度雖然快，但是透過顏色去挑選咖啡生豆會無法將較小的瑕疵分類出來。而只要混入一點瑕疵的咖啡生豆，就會影響最終咖啡的品質與口感。
本研究先建置一個拍攝阿拉比卡咖啡生豆的系統，使用影像處理技術擷取咖啡生豆的影像，將其分類並整合成資料集，用於訓練和驗證卷積神經網路模型。分別使用六種不同的模型(DenseNet121、EfficientNetB0、InceptionV3、MobileNetV2、ResNet50、Xception)進行遷移學習，綜合評估模型的準確率、精確率、召回率、F1-Score及模型的辨識速度，發現ResNet50具有較好的準確率(96%)與快速的辨識速度(0.836s)。最後整合影像處理技術與卷積神經網路模型，對咖啡生豆進行分類。

In modern society, coffee, with its rich historical background, has become an indispensable beverage in the daily lives of most people. Ensuring its consistent flavor is of paramount importance. Common methods for selecting coffee beans include manual selection and color sorter machines. Manual selection is a time-consuming and mentally demanding task, whereas color sorter machines offer speed but may fail to detect smaller defects based solely on color. Even a minor inclusion of defective coffee beans can impact the quality and flavor of the final coffee.
This study initially establishes a system for capturing images of Arabica coffee beans. Image processing techniques are utilized to extract images of the coffee beans, classify them, and create a dataset for training and validating convolutional neural network models. Six different models (DenseNet121, EfficientNetB0, InceptionV3, MobileNetV2, ResNet50, Xception) are employed for transfer learning. The models are comprehensively evaluated based on accuracy, precision, recall, F1-Score, and recognition speed. ResNet50 is found to exhibit superior accuracy (96%) and rapid recognition speed (0.836 seconds). Finally, image processing techniques and convolutional neural network models are integrated to classify coffee beans.

摘要 …………………………………………………………i
Abstract …………………………………………………………ii
誌謝 …………………………………………………………iii
目錄 …………………………………………………………iv
表目錄 …………………………………………………………vii
圖目錄 …………………………………………………………viii
第一章 緒論 …………………………………………………………1
1.1 研究背景 …………………………………………………………1
1.2 研究動機與目的 …………………………………………………………2
第二章 理論與文獻探討 …………………………………………………………3
2.1 光學色彩選別機 …………………………………………………………3
2.2 咖啡生豆的品質定義 …………………………………………………………5
2.2.1 咖啡生豆 …………………………………………………………5
2.2.2 咖啡生豆處理方式 …………………………………………………………5
2.2.3 咖啡生豆分級協議 …………………………………………………………6
2.3 文獻探討 …………………………………………………………10
2.3.1 影像處理相關文獻探討 …………………………………………………………10
2.3.2 深度學習相關文獻探討 …………………………………………………………11
2.4 影像處理技術 …………………………………………………………12
2.4.1 RGB色彩空間 …………………………………………………………12
2.4.2 YUV色彩空間 …………………………………………………………13
2.4.3 灰階色彩轉換 …………………………………………………………13
2.4.4 高斯模糊 …………………………………………………………15
2.4.5 影像二值化 …………………………………………………………16
2.4.6 膨脹（Dilation） …………………………………………………………17
2.4.7 侵蝕（Erosion） …………………………………………………………18
2.4.8 斷開（Opening） …………………………………………………………19
2.4.9 閉合（Closing） …………………………………………………………20
2.4.10 影像輪廓偵測 …………………………………………………………21
2.5 類神經網路 …………………………………………………………22
2.5.1 類神經元數學模型 …………………………………………………………22
2.5.2 類神經網路架構 …………………………………………………………24
2.6 卷積神經網路 …………………………………………………………24
2.6.1 卷積層 …………………………………………………………25
2.6.2 池化層 …………………………………………………………25
2.6.3 全連接層 …………………………………………………………26
2.6.4 深度可分離式卷積 …………………………………………………………26
第三章 實驗步驟 …………………………………………………………29
3.1 研究設備 …………………………………………………………29
3.1.1 工業相機 …………………………………………………………29
3.1.2 環形光源 …………………………………………………………30
3.1.3 LattePanda Delta …………………………………………………………31
3.1.4 英特爾®神經計算棒2 …………………………………………………………31
3.1.5 進料架構 …………………………………………………………32
3.1.6 分離型調速馬達 …………………………………………………………34
3.1.7 硬體架構 …………………………………………………………35
3.1.8 軟體設備 …………………………………………………………36
3.2 實驗流程 …………………………………………………………37
3.2.1 拍攝咖啡生豆資料集 …………………………………………………………37
3.2.2 訓練卷積神經網路模型 …………………………………………………………38
3.2.3 咖啡生豆影像辨識與分類 …………………………………………………………39
3.2.4 混淆矩陣 …………………………………………………………40
3.2.5 效能衡量指標 …………………………………………………………40
3.2.6 ROC曲線法 …………………………………………………………41
3.2.7 AUC …………………………………………………………42
第四章 結果與討論 …………………………………………………………43
4.1 咖啡生豆資料集 …………………………………………………………43
4.2 卷積神經網路模型訓練結果 …………………………………………………………44
4.3 效能衡量指標 …………………………………………………………45
4.3.1 DenseNet121預測結果 …………………………………………………………45
4.3.2 EfficientNetB0預測結果 …………………………………………………………46
4.3.3 InceptionV3預測結果 …………………………………………………………47
4.3.4 MobileNetV2預測結果 …………………………………………………………48
4.3.5 ResNet50預測結果 …………………………………………………………49
4.3.6 Xception預測結果 …………………………………………………………50
4.3.7 所有模型的平均效能衡量指標 …………………………………………………………51
4.4 ROC曲線 …………………………………………………………52
4.5 模型相關參數比較 …………………………………………………………53
第五章 結論 …………………………………………………………54
第六章 未來展望 …………………………………………………………55
參考文獻 56
EXTENDED ABSTRACT …………………………………………………………59
1. INTRODUCTION …………………………………………………………60
2. EXPERIMENTAL …………………………………………………………61
3. RESULTS AND DISCUSSION …………………………………………………………61
4. CONCLUSIONS …………………………………………………………63
REFERENCES …………………………………………………………64

[1]財政部統計處, 97-107 年咖啡豆進口量成長 2.1 倍，今年進口量規模可望接續 107 年再創新高. 財政統計通報第22號, 2019.
[2]Organization, I.C., World coffee consumption. 2021.
[3]Vincent, J., Green coffee processing, in Coffee: Volume 2: Technology. 1987, Springer. p. 1-33.
[4]Ghosh, P. and N.J.I.J.E.R.T. Venkatachalapathy, Processing and drying of coffee–a review. 2014. 3(12): p. 784-794.
[5]Haile, M., W.H.J.C.-P. Kang, and Research, The harvest and post-harvest management practices’ impact on coffee quality. 2019: p. 1-18.
[6]PUERTA, G.I., Influencia del proceso de beneficio en la calidad del café. 1999.
[7]Lee, D.-J. and R.S. Anbalagan. High-speed automated color-sorting vision system. in Optical Engineering Midwest'95. 1995. SPIE.
[8]Boyd Jr, A., Potential applications of electric color sorting techniques in seed technology. 2021.
[9]Pearson, T., et al., Color image based sorter for separating red and white wheat. 2008. 2: p. 280-288.
[10]He, L., Z. Niu, and D. Chen. Research on development of color sorter using triz. in Knowledge Enterprise: Intelligent Strategies in Product Design, Manufacturing, and Management: Proceedings of PROLAMAT 2006, IFIP TC5 International Conference, June 15–17, 2006, Shanghai, China. 2006. Springer.
[11]Kosalos, J., et al., Arabica Green Coffee: Defect Handbook. 2013: Specialty Coffee Association of America.
[12]林晋億, 全彩機械視覺於咖啡生豆檢測之研究, in 電子與光電工程研究所碩士班. 2013, 國立雲林科技大學: 雲林縣. p. 66.
[13]凃亞廷, 咖啡公豆母豆自動選別裝置之開發, in 生物產業機電工程學系所. 2014, 國立中興大學: 台中市. p. 113.
[14]Mebatsion, H., et al., Automatic classification of non-touching cereal grains in digital images using limited morphological and color features. 2013. 90: p. 99-105.
[15]鄭志祥, 影像次像素應用在米粒檢測之研究, in 農業機械工程學系. 2002, 國立中興大學: 台中市. p. 92.
[16]葉翊暉, 發展基於深度學習之咖啡生豆顏色分類系統, in 機械工程研究所. 2021, 中原大學: 桃園縣. p. 47.
[17]Unal, Y., et al., Application of pre-trained deep convolutional neural networks for coffee beans species detection. 2022. 15(12): p. 3232-3243.
[18]Podpora, M., G.P. Korbas, and A. Kawala-Janik. YUV vs RGB-Choosing a Color Space for Human-Machine Interaction. in FedCSIS (Position Papers). 2014. Citeseer.
[19]Bovik, A.C., The essential guide to image processing. 2009: Academic Press.
[20]Haralick, R.M., et al., Image analysis using mathematical morphology. 1987(4): p. 532-550.
[21]Duda, R.O. and P.E.J.C.o.t.A. Hart, Use of the Hough transformation to detect lines and curves in pictures. 1972. 15(1): p. 11-15.
[22]Canny, J.J.I.T.o.p.a. and m. intelligence, A computational approach to edge detection. 1986(6): p. 679-698.
[23]Vincent, O.R. and O. Folorunso. A descriptive algorithm for sobel image edge detection. in Proceedings of informing science & IT education conference (InSITE). 2009.
[24]Suzuki, S.J.C.v., graphics, and i. processing, Topological structural analysis of digitized binary images by border following. 1985. 30(1): p. 32-46.
[25]Jain, A.K., J. Mao, and K.M.J.C. Mohiuddin, Artificial neural networks: A tutorial. 1996. 29(3): p. 31-44.
[26]Hecht-Nielsen, R., Theory of the backpropagation neural network, in Neural networks for perception. 1992, Elsevier. p. 65-93.
[27]Sharma, S., S. Sharma, and A.J.T.D.S. Athaiya, Activation functions in neural networks. 2017. 6(12): p. 310-316.
[28]Sanger, T.D.J.N.n., Optimal unsupervised learning in a single-layer linear feedforward neural network. 1989. 2(6): p. 459-473.
[29]Albawi, S., T.A. Mohammed, and S. Al-Zawi. Understanding of a convolutional neural network. in 2017 international conference on engineering and technology (ICET). 2017. Ieee.
[30]Hubel, D.H. and T.N.J.T.J.o.p. Wiesel, Receptive fields and functional architecture of monkey striate cortex. 1968. 195(1): p. 215-243.
[31]Fukushima, K.J.B.c., Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. 1980. 36(4): p. 193-202.
[32]LeCun, Y., et al., Gradient-based learning applied to document recognition. 1998. 86(11): p. 2278-2324.
[33]Lin, M., Q. Chen, and S.J.a.p.a. Yan, Network in network. 2013.
[34]Chollet, F. Xception: Deep learning with depthwise separable convolutions. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
[35]Huang, G., et al. Densely connected convolutional networks. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
[36]Tan, M. and Q. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. in International conference on machine learning. 2019. PMLR.
[37]Szegedy, C., et al. Rethinking the inception architecture for computer vision. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
[38]Sandler, M., et al. Mobilenetv2: Inverted residuals and linear bottlenecks. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.
[39]He, K., et al. Deep residual learning for image recognition. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
[40]Luque, A., et al., The impact of class imbalance in classification performance metrics based on the binary confusion matrix. 2019. 91: p. 216-231.
[41]Fawcett, T.J.P.r.l., An introduction to ROC analysis. 2006. 27(8): p. 861-874.
[42]Bradley, A.P.J.P.r., The use of the area under the ROC curve in the evaluation of machine learning algorithms. 1997. 30(7): p. 1145-1159.