臺灣博碩士論文加值系統

本論文使用YOLO(You Only Look Once)-V3、IOU(Intersection over Union)追蹤以及CNN(Convolutional Neural Network)等人工智慧演算法提出一個可辨識水果外在品質的系統。本系統主要是以YOLO演算法進行水果偵測與IOU追蹤演算法對水果持續追蹤，在追蹤的過程中辨識水果是否有損壞的部分，並挑選出來。本系統使用Jetson TX2 嵌入式開發平台作為載台，並且利用STM32處理器控制輸送台的閘道開關。
透過本系統可針對以小型且圓形類的水果進行篩選，本系統提供一個有效的開發流程，提高開發效率，並設計一圖形化介面程式，用來收取數據、評估模型，並且掌握整個系統運作的過程。在實驗部分，測試六類水果，在水果測試照片總數為4500張做為實驗測試，其辨識結果可達約88%的準確度，而在單顆偵測上平均精度均值(mean Average Precision, mAP )為75%。

The thesis presents a proposed system that uses YOLO (You Only Look Once)-V3 algorithm, IOU (Intersection over Union) tracking method, and CNN (Convolutional Neural Network) classifier to identify the external quality of fruits. The system mainly uses the YOLO-V3 algorithm to perform the fruit detection process, uses the IOU tracking algorithm to track the designated fruits continuously, and identifies fruits during the tracking processes. It can pick up good fruits through controlling the switched gap of conveying platform. It performs the software programs on the Jetson TX2 embedded development platform and uses the STM32 processor to control the switched gap.
The proposed system can detect small and round fruits under an effective development process. To improve the efficiency of system, a graphic user interface is also designed to control , collect data, evaluate models,and monitor the entire system operation. The experimental results show that our proposed system can achieve up to 88% of the accuracy rate, 75% of the mean Average Precision (mAP) after testing 4,500 images of fruits.

中文摘要 I
英文摘要 II
目 錄 IV
表目錄 VII
圖目錄 VIII
一、 緒論 1
1.1 動機與目的 1
1.2 研究工具介紹 3
二、 文獻探討與相關研究 7
2.1 卷積神經網路 7
2.2 物件偵測相關文獻與應用 12
2.3 影像辨識之方法 15
2.4 追蹤相關文獻與應用 16
2.5 水果篩選設備 17
2.6 神經網路開發工具 18
三、 系統架構與實作方法 19
3.1 系統架構(SYSTEM ARCHITECTURE) 19
3.1.1系統介紹 19
3.2 硬體架構 22
3.3 軟體架構 27
3.3.1 前方閘道器 28
3.3.2 TINY-YOLO物件偵測模型 28
3.3.3 比較消除演算法 30
3.3.4 追蹤演算法 31
3.3.5 水果分類辨識模型 32
3.3.6 UFF模型 33
3.3.7 後方閘道器 36
3.3.9 物件偵測模型 37
3.3.8圖形化介面視窗 46
四、 實驗結果 50
4.1 辨識模型評估與訓練(數據集一) 50
4.1.1模型訓練 51
4.1.2 交叉驗證K-FOLD 52
4.1.3 混淆矩陣（CONFUSION MATRIX） 54
4.2 物件偵測模型評估與訓練(數據集二) 57
4.2.1 PR曲線(PRECISION-RECALL) 58
4.2.2 平均精度均值(MEAN AVERAGE PRECISION),MAP 63
4.2.3 模型訓練 67
4.2.4 結果圖(數據集一、二) 69
4.3 辨識模型評估與訓練(數據集三) 70
4.3.1模型訓練 70
4.3.2PR曲線(PRECISION-RECALL) 71
4.3.3平均精度均值MEAN AVERAGE PRECISION(MAP) 72
4.3.4結果圖(數據集三) 73
4.4 平台實際測試 74
4.5 伺服馬達訊號圖 75
五、 結論與未來展望 77
5.1 結論 77
5.2 未來展望 77
六、 參考文獻 78

[1]國家發展委員會，2018，「中華民國人口推估（2018至2065年）」報告。 Available at https://pop-proj.ndc.gov.tw/download.aspx?uid=70&pid=70
[2]行政院農業委員會農糧署，2012，蔬果品質分級標準暨包裝規格手冊-水果篇。Available at https://www.afa.gov.tw/ebook/afa/ebook13_1/#p=14
[3]A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam,”Mobilenets: Efficient convolutional neural networks for mobile vision applications,” arXiv preprint arXiv:1704.04861, 2017.
[4]M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4510-4520, 2018.
[5]Y. Chen, H. Fang, B. Xu, Z. Yan, Y. Kalantidis, M. Rohrbach, S. Yan, and J. Feng, 2019,”Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution,” arXiv preprint arXiv:1904.05049, 2019.
[6]F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. J. a. p. a. Keutzer, ”SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size,” arXiv preprint arXiv:1602.07360, 2016.
[7]W. Liu, D. Anguelov, D. Erhan, C. Szegedy, and S. Reed, 2016, “Ssd: Single shot multibox detector,” Proceedings of European conference on computer vision, Springer, pp. 21-37, 2016.
[8]J. Redmon and A. Farhadi, 2018, ”Yolov3: An incremental improvement, “ arXiv preprint arXiv:1804.02767, 2018.
[9]M. Apte, S. Mangat, and P. Sekhar, “YOLO Net on iOS, “cs231n.stanford.edu, 2017.
[10]A. Womg, M. J. Shafiee, F. Li, and B. Chwyl, “Tiny ssd: A tiny single-shot detection deep convolutional neural network for real-time embedded object detection,” Proceedings of 2018 15th IEEE Conference on Computer and Robot Vision (CRV) , pp. 95-101, 2018.
[11]T. Santad, P. Silapasupphakornwong, W. Choensawat, and K. Sookhanaphibarn, “Application of YOLO Deep Learning Model for Real Time Abandoned Baggage Detection,” Proceedings of 2018 IEEE 7th Global Conference on Consumer Electronics (GCCE), pp. 157-158, 2018.
[12]Yogesh, A. K. Dubey, “Fruit defect detection based on speeded up robust feature technique,” Proceedins of 2016 5th IEEE International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions)(ICRITO) , pp. 590-594, 2017.
[13]Z. M. Khaing, Y. Naung, and P. H. Htut, “Development of control system for fruit classification based on convolutional neural network,” Proceedings of 2018 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (EIConRus), pp. 1805-1807, 2018.
[14]Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, 1998, ”Gradient-based learning applied to document recognition,” Proceedings of the IEEE, Vol. 86, No. 11, pp. 2278-2324, 1998.
[15]D. Jung, J.-W. Son, and S.-J. Kim, ”Shot category detection based on object detection using convolutional neural networks,” Proceedings of 2018 20th IEEE International Conference on Advanced Communication Technology (ICACT), pp. 36-39, 2018.
[16]E. Bochinski, V. Eiselein, and T. Sikora, “High-speed tracking-by-detection without using image information,” Proceedings of 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), pp. 1-6, 2017.
[17]H. Ren, F. Xu, F. Zou, K. Jia, P. Di, and J. Kang, 2018, “Multi-pedestrian Tracking Based on Social Forces,” Proceedings of 2018 IEEE International Conference on Intelligence and Safety for Robotics (ISR), pp. 527-532, 2018.
[18]S. Liu, X. Li, M. Gao, Y. Cai, R. Nian,P. Li, Y. Tianhong and A. Lendasse, 2018, ”Embedded Online Fish Detection and Tracking System via YOLOv3 and Parallel Correlation Filter,” Proceedings of IEEE OCEANS 2018 MTS/IEEE Charleston, pp. 1-6, 2018.
[19]台州斯帕克儀器儀表儀錶有限公司，2019，供應果蔬分選機水果大小分選機。 Available at https://big5.made-in-china.com/gongying/spk666-YomQUAWGqFhc.html.
[20]T. S. consortium, 2019, greenhouse harvesting robots. Available at http://www.sweeper-robot.eu/.
[21]J. Hale, 2019, Which Deep Learning Framework is Growing Fastest. Available at https://towardsdatascience.com/which-deep-learning-framework-is-growing-fastest-3f77f14aa318
[22]A Keras implementation of YOLOv3, 2018. Available at https://github.com/qqwweee/keras-yolo3.
[23]G. Chen, P. Chen, Y. Shi, C.-Y. Hsieh, B. Liao, and S. Zhang, “Rethinking the Usage of Batch Normalization and Dropout in the Training of Deep Neural Networks,” arXiv preprint arXiv:1905.05928, 2019.
[24]J. Pedoeem and R. Huang, “Yolo-lite: A real-time object detection algorithm optimized for non-gpu computers,” arXiv preprint arXiv:1811.05588, 2018.
[25]Object-Detection-Metrics, 2018. Available at https://github.com/rafaelpadilla/Object-Detection-Metrics.
[26]Image Net, Stanford Vision Lab, Princeton University, 2016. Available at http://www.image-net.org/.