臺灣博碩士論文加值系統

菇類產業目前是台灣智慧農業的重要產業之一，僅管台灣農業培養技術純熟，但是以目前農業的自動化程度與其他先進國家相對較低。本研究主要目的在利用深度學習影像辨識方法，檢測菇類太空包製包自動化生產線之瑕疵品。本研究利用機械手臂結合三維視覺模組，並搭配第七軸實現機械手臂多工的系統，及深度學習方法來實現產線中瑕疵品之自動化檢測。
本文利用投射LED光源之結構光方式投射至待測物件，使用結構光之投影方法來有效的辨識大範圍且複雜之物件。搭配雙鏡頭CMOS工業攝影機擷取圖像，並藉由軟體計算建立物件之3D點雲圖，能夠獲得物件的詳細資訊，建立太空包之圖庫。藉由標籤訓練集進行深度學習訓練來達成影像辨識，本文使用相同訓練參數進行個別訓練，並將影像資料分為四組圖資。系統透過圖形處理器並利用CUDA架構來訓練深度學習模型，節省訓練時間並可提升物件辨識率。研究中使用可變夾爪結合夾治具以確實的夾取太空包。藉由可變夾爪可調節夾持力與開合度的特性，將夾爪夾持力維持在20N之操作條件，不致於破壞太空包之結構與內部的孔洞使其無法重複使用。由實驗結果顯示，利用深度學習方法採用圖資1的影像數量以及在200lm至800lm的環境光源中，擁有高達0.989的平均辨識信心。並且利用可變夾爪進行30次重複夾取實驗，仍可維持太空包內孔徑在標準範圍25mm以上，符合菇類太空包製包的標準。

At present, The mushroom industry is still one of Taiwan's important industries. Although Taiwan's agricultural cultivation technology is skillful, the degree of automation of agriculture is relatively low compared with other advanced countries.The main purpose of the research is to use the image recognition method of Deep Learning algorithm to solve the removal of products on the automatic production line of mushroom PP-bag. This study uses the robotic arm (Robot) combined with visual module to realize the elimination of defective products on the production line by using the depth learning method, and to implement the mechanical arm multitask system with the Seventh Axis.
Because of the complexity of the object range and condition of identification, we use the structure light of the projection led light source to project to the object to be tested, and use the structure light projection method to effectively identify the large range and complex object, combine with the dual-lens CMOS industrial camera to retrieve the image, by software computing to create a 3D point cloud image of the object, we can get detailed information about the object and build a image database of mushroom PP-bag. The deep learning training is performed by the Labeled training dataset to achieve image recognition. The system uses the GPU and uses the CUDA architecture to train the deep learning model, saving training time and improving object recognition rate. Using adaptive robot gripper combined with fixture to indeed capture the mushroom PP-bag. By changing the clamping force and the characteristics of the opening and closing, adjust the best clamping posture, not destroy the space package structure and internal holes so that it can not be reused.The experimental results show that the depth learning algorithm has an average recognition confidence of up to 0.989 in the number of images and the stable ambient light source. Adaptive robot gripper and fixtures is able to accurately capture PP bags even though repeated gripping，The aperture in the same mushroom PP-bag can also be maintained in the standard range of 25mm or more without excessive deformation.

中文摘要 ............................... i
英文摘要 ............................... ii
誌謝 ............................... iv
目錄 ............................... v
表目錄 ............................... vii
圖目錄 ............................... viii
第一章 緒論............................ 1
1.1 前言........................... 1
1.2 研究動機與目的.................. 1
1.3 研究架構........................ 3
第二章 文獻回顧........................ 4
2.1 菇類栽培相關影像辨識回顧......... 4
2.2 結構光原理...................... 6
2.2.1 結構光數學模式................ 6
2.2.2 結構光模式.................... 9
2.2.3 結構光種類.................... 10
2.3 相機校正........................ 10
2.4 機械手臂運動學.................. 14
2.4.1 D-H轉換矩陣................... 14
2.4.2 順向運動學.................... 15
第三章 實驗方法與步驟.................. 18
3.1 結構光影像辨識模組.............. 18
3.1.1 CMOS工業相機................. 18
3.1.2 深度學習基本原理.............. 19
3.1.3 CUDA浮點運算之影像深度學習..... 20
3.1.4 鏡頭影像校正.................. 21
3.1.5 影像擷取介面.................. 22
3.2 電控系統........................ 22
3.2.1 UR5協作型六軸機械手臂.......... 23
3.2.2 Robotiq可變夾爪............... 24
3.2.3 夾治具研製.................... 25
3.2.4 視覺模組與機械手臂位置校正...... 27
3.3 第七軸平台...................... 28
第四章 結果與討論....................... 29
4.1 實驗環境........................ 29
4.1.1 實驗對象...................... 29
4.1.2 實驗場景...................... 30
4.1.3 系統架構...................... 32
4.2 實驗流程........................ 33
4.3 影像辨識實驗結果................. 34
4.4 太空包夾取與重複性實驗結果........ 37
第五章 結論與未來展望................... 40
5.1 結論............................ 40
5.2 未來展望........................ 40
參考文獻 ............................... 41
附錄一 ................................. 44
英文論文大綱

[1]DIGITIMES企劃 (2014)。自動化到智動化 製造業機器人應用趨勢。民107年6月21日，取自https://www.digitimes.com.tw/iot/article.asp?cat=130&id=0000372926_YELLGRJS1A8IWZ77PEHX2
[2]L. Gudrun (2017), “How Robots Conquer Industry Worldwide,” International Federation of Robotics IFR.
[3]R. K. Naresh, L. Seung-Yeol, J. Hee-Young, J. Mansik, & K. Jeehyun, “Optical inspection and monitoring of moisture content in Pleurotus eryngii during storage life by refrigeration (2017),” International Journal of Applied Engineering Research, Vol 12, pp.5011-5015.
[4]M. Alireza (2013), “Computer Vision Algorithms for an Automated Harvester,” The University of Western Ontario.
[5]J. N. Reed, S. J. Miles, J. Butler, M. Baldwn, & R. Noble (2001), “Automatic Mushroom Harvester Development,” Journal of Agricultural Engineering Research, Vol 78, pp. 15-23.
[6]R. D. Tillett, & B. G. Batchelor (1991), “An Algorithm for Locating Mushrooms in a Growing Bed,” Computers and Electronics in Agriculture, Vol 6, pp. 191-200.
[7]D. Hongshe, S. Jinguo, & G. Qin (2010), “A Fruit Size Detecting and Grading System Based on Image Processing,” IEEE Second International Conference on Intelligent Human-Machine Systems and Cybernetics.
[8]I. D. Reid (1996), “Projective Calibration of a Laser-Stripe Range Finder,” Image and Vision Computing, Vol 14, pp. 659-666.
[9]W. Shiuanshiau (2005), “Laser Structure Light Applies to Measure the Surface and Three-Dimensional Structure of Objects Meticulously,” National Central University Department of Optics and Photonics.
[10]韓軍、劉鈞 (2012)。工程光學（初版）。北京：國防工業。
[11]MathWorks, “What Is Camera Calibration?,” Retrieved June 23, 2018, from
http://www.mathworks.com/help/vision/ug/camera-calibration.html.
[12]OpenCV, “Camera Calibration and 3D Reconstruction,” Retrieved June 23, 2018, from
https://docs.opencv.org/2.4/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html?highlight=calibratecamera.
[13]B. Roland, K. Conrad, F. Tilo, & E. Norbert (2011), “Kinematics Analysis of a 3-DOF Joint for a Novel Hyper-Redundant Robot Arm,” IEEE International Conference on Robotics and Automation, pp. 3224-3229.
[14]H. Han-Pang, C. Yen-Tsung, W. Ren-Jeng, & C. Meng-Ku (2010), “Development of the Multi-Axis Control Platform for Robot Arm,” IEEE Proceedings of SICE Annual Conference, pp. 2326-2331.
[15]Z. Zhang (2000), “A Flexible New Technique for Camera Calibration,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, pp.1330-1334.
[16]H. Richard (1999), “Theory and Practice of Projective Rectification,” International Journal of Computer Vision, Vol. 35, pp 115-127.
[17]F. F. Hafidz, & S. J. Muhd (2008), “Modeling and Simulation of Programmable Universal Machine for Assembly (PUMA) industrial robot for automotive-related assembly process,” IEEE International Symposium on Information Technology.
[18]黃緒哲（民77年）。基本機器人學（初版）。臺北市：徐氏文教。
[19]J. Denavit, & R. S. Hartenberg (1955), “A Kinematic Notation for Lower-Pair Mechanisms Based on Matrices,” ASME Journal of Appled Mechanics, Vol. 22 pp. 215-221.
[20]Y. Fu-wen (2012), “Mechanical Arm Control Combined with Image System,” National Taiwan Normal University Department of Mechatronic Engineering.
[21]R. P. Paul (1981), “Robot Manipulators: Mathematics, Programming, and Control,” MIT Press, Cambridge, Mass., pp. 119-155.
[22]江修、黃偉峰（2006）。六軸機械臂之控制理論分析與應用【專刊】。機械工業，277，57-73。
[23]H. Kelsey (2013), “Analytic Inverse Kinematics for the Universal Robots UR-5/UR-10 Arms,”.
[24]TOSHIBA, “CMOS Camera BU Series Specifications,”.
[25]D. Li, & Y. Dong (2013), “Deep Learning: Methods and Applications,” Foundations and Trends in Signal Processing, Vol. 7, pp. 197-387.
[26]MathWorks, “What is Deep Learning?,” Retrieved June 25, 2018, from
https://ww2.mathworks.cn/en/discovery/deep-learning.html?s_tid=srchtitle
[27]S. Juergen (2015), “Deep Learning in Neural Networks: an Overview,” Neural Networks, Vol. 61, pp. 85-117.
[28]NVIDIA (2018), “CUDA C Programming Guide,” PG-02829-001_v9.2.88.
[29]NVIDIA（2015）。Geforce GTX950規格。民107年6月25日，取自
https://www.geforce.com.tw/hardware/desktop-gpus/geforce-gtx-950/specifications.
[30]J. E. Colgate, W. Wannasuphoprasit, & M. A. Peshkin (1996), “Cobots: Robots for Collaboration with Human Operators,” Proceedings of the International Mechanical Engineering Congress and Exhibition, Atlanta, GA, DSC-Vol. 58, pp.433-39.
[31]Universal Robots, “UR5 Technical details,” Retrieved June 25, 2018, from
https://www.universal-robots.com/media/1801303/eng_199901_ur5_tech_spec_web_a4.pdf.
[32]Robotiq (2016), “Adaptive Robot Gripper 2-Finger (85 & 140) Instruction Manual,” Robotiq Inc.
[33]S. M. Wang, & K. F. Ehmann (1999), “Measurement Methods for the Position Errors of a Multi-Axis Machine. Part 1: Principles and Sensitivity Analysis,” International Journal of Machine Tools and Manufacture, Vol. 39, pp. 951-964.