臺灣博碩士論文加值系統

近年來，菇類生產已成為台灣農業的重點產業之一，但大部分作業仍高度依賴人力。以杏鮑菇(Pleurotus eryngii)的分級為例，主要透過判斷菇體外觀來進行分級。然而，人員的培訓非常耗時，且可能會因為生理和心理因素，導致不適當的分級。
本研究提出以機器視覺結合深度學習之卷積神經網路為基礎，於樹莓派上研究與開發一套杏鮑菇自動化分級系統；初期用於拍攝杏鮑菇圖像並建立圖像集，部分圖像用於卷積神經網路的訓練和驗證，其他則用於神經網路的泛化能力測試。
本研究開發的杏鮑菇自動化分級系統在測試集中擁有93.18%的準確度，而在多項模型能力評估指標中皆有傑出的評比表現。顯示本研究開發的杏鮑菇自動化分級系統具有良好的泛化及分類能力，可以有效、準確地對杏鮑菇進行分級。

In recent years, mushroom production has become one of the important industries in Taiwan, but most of the operations still highly rely on manpower. Taking the classification of Pleurotus eryngii for an example, it is mostly classified by judging its appearance.
This study developed an automated classification system of Pleurotus eryngii based on machine vision and convolutional neural network (CNN). In this study, we designed the system based on raspberry pi, which used to collect Pleurotus eryngii image samples. Some of the samples were used for training and validation of CNN. Other samples were used to test the generalization ability of neural network model.
The automated classification system of Pleurotus eryngii developed in this study has an accuracy of 93.18% in the final test. The system also has outstanding performance in multiple model evaluation indicators. Obviously, the system has a perfect generalization ability on the neural network model, and it can effectively and accurately classify Pleurotus eryngii.

摘要.............................................................i
Abstract........................................................ii
誌謝.............................................................iii
目錄.............................................................iv
表目錄...........................................................vi
圖目錄...........................................................vii
第一章 緒論..................................................1
1.1 研究背景.................................................1
1.2 研究動機與目的.......................................3
1.3 論文架構.................................................4
1.4 研究流程.................................................4
第二章 文獻回顧與方法.................................5
2.1 深度學習.................................................5
2.2 卷積神經網路............................................5
2.2.1 卷積層(Convolutional Layer).................6
2.2.2 池化層(Pooling Layer)...........................7
2.2.3 全連接層(Fully-Connected Layer).........7
2.2.4 激活函數(Activation Function)..............8
2.2.5 批標準化(Batch Normalization).............9
2.2.6 優化器(Optimizer)................................10
2.3 模型評估方法.........................................11
2.3.1 過擬合(Overfitting)..............................11
2.3.2 混淆矩陣(Confusion Matrix).................12
2.3.3 ROC曲線與PR曲線.............................13
2.3.4 Kappa一致性係數...............................14
第三章 研究材料與步驟................................15
3.1 輸送分級平台..........................................15
3.1.1 動作流程.............................................15
3.1.2 硬體設計..............................................16
3.1.3 攝像模組比較........................................21
3.1.4 控制電路設計........................................22
3.1.5 程式設計..............................................23
3.2 杏鮑菇圖像資料集建立.............................24
3.2.1 實驗場域杏鮑菇分級流程與概況............24
3.2.3 杏鮑菇圖像拍攝...................................25
第四章 研究結果與討論.................................26
4.1 資料集切割............................................26
4.2 深度學習設備與環境..............................27
4.3 神經網路架構..........................................28
4.4 模型訓練成果與成效評估........................29
4.4.1 學習曲線.............................................29
4.4.2 模型分類結果......................................30
4.4.3 混淆矩陣.............................................31
4.4.4 ROC曲線與PR曲線.............................32
4.4.5 Kappa一致性係數...............................33
第五章 結論與未來展望................................34
5.1 結論......................................................34
5.2 未來展望................................................35
參考文獻.....................................................36
Extended Abstract.......................................38

[1] 石信德等 (2019)。菇類智慧化生產之現況與展望。菇類智慧化生產與農場經營管
理研討會專刊，台中，行政院農業委員會農業試驗所，頁29-34。
[2] 周榮源、石信德 (2018)。智慧化菇類生產技術之前瞻與未來發展。第27屆近代
工程技術討論會，台北。
[3] 內政部統計處 (2018)。107年第15週內政部統計通報_高齡社會。
[4] 行政院主計總處 (2018)。人力資源調查統計年報，頁16。
[5] 嘉義縣東石合作農場 (2019)。好菇菇小幫手。檢自
https://tscf.myorganic.org.tw/%E5%85%A8%E7%94%A2%E5%93%81/GOOD-GOOD-%E5%A5%BD%E8%8F%87 (2019年1月22日)。
[6] 正郁農場。杏鮑菇簡介。檢自
http://www.focusline.com.tw/vegetable/summary.aspx (2019年1月22日)。
[7] Wikipedia. (2019). Deep Learning. Retrieved from
https://en.wikipedia.org/wiki/Deep_learning (March 12,2019).
[8] Afshine, A., Shervine, A. Deep Learning cheatsheets for Stanford's CS 230. Retrieved from https://stanford.edu/~shervine/teaching/cs-230/cheatsheet-convolutional-neural-networks (May 12,2019).
[9] 李宏毅 (2017)。Machine Learning (2017,Fall)。檢自http://speech.ee.ntu.edu.tw/~tlkagk/courses_ML17_2.html (2019年3月5日)。
[10] 莫煩Python (2018)。激勵函數(Activation Function)。檢自
https://morvanzhou.github.io/tutorials/machine-learning/ML-intro/3-04-activation-function/ (2019年4月19日)。
[11] Clevert, D., Unterthiner, T., & Hochreiter, S. (2016). Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs). CoRR, abs/1511.07289.
[12] Ioffe, S., & Szegedy, C. (2015). Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. ArXiv, abs/1502.03167.
[13] Jeremy Jordan. (2019). Setting the learning rate of your neural network. Retrieved from https://www.jeremyjordan.me/nn-learning-rate/ (March 7,2019).
[14] Kingma, D.P., & Ba, J. (2015). Adam: A Method for Stochastic Optimization. CoRR, abs/1412.6980.
[15] Dozat, T. (2016). Incorporating Nesterov Momentum into Adam. International Conference on Learning Representations (ICLR),2016.
[16] AWS (2019)。模型擬合：低度擬合與過度擬合。檢自https://docs.aws.amazon.com/zh_tw/machine-learning/latest/dg/model-fit-underfitting-vs-overfitting.html (2019年3月12日)。
[17] Wikipedia. (2019). Confusion matrix. Retrieved from
https://en.wikipedia.org/wiki/Confusion_matrix (May 14,2019).

[18] Yang, S., & Berdine, G. (2017). The receiver operating characteristic (ROC) curve. The Southwest Respiratory and Critical Care Chronicles, 5(19), pp. 34-36.
[19] Fawcett, T. (2006). An introduction to ROC analysis. Pattern Recognition Letters, 27, pp.861-874.
[20] Cohen, J. (1960). A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1), pp.37–46.
[21] Viera, A.J., & Garrett, J.M. (2005). Understanding interobserver agreement: the kappa statistic. Family medicine, 37 5, 360-3.
[22] Wikipedia. (2019). Raspberry Pi. Retrieved from https://en.wikipedia.org/wiki/Raspberry_Pi (Jan 15,2019).
[23] Raspberry Pi Pinout. Retrieved from https://pinout.xyz/ (Jan 15,2019.)
[24] Raspberry Pi Camera Module v2. Retrieved from https://www.raspberrypi.org/products/camera-module-v2/ (Jan 15,2019).
[25] Dave Jones. (2019). Picamera Documentation. Retrieved from https://picamera.readthedocs.io/en/latest/ (Jan 15,2019).
[26] Panasonic. Ultra-slim Photoelectric Sensor EX-10 Ver.2. Retrieved from https://www3.panasonic.biz/ac/ae/fasys/sensor/photoelectric/ex-10/index.jsp (Dec 6,2018).
[27] 東方馬達 (2018)。步進馬達大小事5分鐘就上手。檢自 https://www.orientalmotor.com.tw/om/knowledge/uroko_st/st02.html (2019年1月17日)。
[28] 斯泰普科技 (2013)。42系列两相混合式步进电机。檢自 http://www.sitaipu.com/cpzx/bjdjxl/lxbjdj/68.html (2019年1月17日)。
[29] DFROBOT，TB6600 步进电机驱动器用户手册，檢自 http://www.dfrobot.com.cn/images/upload/File/DRI0043/20161012102520yg1obn.pdf (2019年1月17日)。
[30] Canon。測光模式全解讀。檢自https://www.canon.com.hk/tc/club/article/itemDetail.do?itemId=10437 (2018年12月17日)。
[31] TensorFlow. Retrieved from https://www.tensorflow.org/ (Nov 6,2018).
[32] Keras. Retrieved from https://keras.io/ (Nov 6,2018).
[33] Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision (pp. 618-626).