臺灣博碩士論文加值系統

摘要
本篇論文主要提出的研究動機是尋找基隆河的主要汙染源，基於這項動機，我們著重於每月多次的定點檢視河川的污染物，再透過自主無人機所結合YOLO的部份構成的影像辨識系統，於固定的時間與地點對廢棄物與垃圾進行辨識，並將結果上傳至資料庫進行比對分析，並同步給政府相關權責單位，協助政府單位分析基隆河汙染源頭的重點區域。本研究主要是利用不同YOLO的架構於河岸垃圾的辨識，另外，比較不同超參數設定於影像辨識的效果。

Abstract
The primary research motivation of this paper is to find the main pollution source of the Keelung River. Based on this motivation, we focus on the fixed-position inspection of the pollutants on the riverside several times a month and then combine the part of YOLOs with the autonomous drone. The image recognition system identifies waste and garbage at fixed places biweekly and uploads the results to the database for comparison and analysis. The data is then sent to relevant government authorities to assist government units in analyzing critical areas of the pollution source of the Keelung River. This study applies different YOLO structures to riverside garbage identification. In addition, different hyperparameter settings are used to compare its efficiency in image identification.

Contents
摘要 I
Abstract II
誌謝 III
Contents IV
List of Figures VI
List of Tables IX
Chapter 1: Introduction 1
1.1 Preface 1
1.2 Research Motivation and Goal 1
1.3 Literature Review 2
1.4 Proposed Method 3
Chapter 2: System Design 4
2.1 UAV System 4
2.2 Power System 7
Chapter 3: Image-Based Trash Detector 9
3.1 Review of YOLO_CNN [22] 9
3.2 Object Detection 9
3.2.1 The Progress of YOLOv4 [5] 9
3.2.2 The Progress of YOLOv5 [10] 13
3.2.3 The Progress of YOLOv7 [11] [12] 16
3.3 Keelung-River Litter Dataset 34
Chapter 4: Experiments and Results 47
4.1 Trash Detector Real-World Testing by YOLOs (based on Keelung River dataset) 47
4.1.1 Trash Detector Real-World Testing by YOLOv4 (based on Keelung River dataset) 47
4.1.2 Trash Detector Real-World Testing by YOLOv5 (based on Keelung River dataset) 56
4.1.3 Trash Detector Real-World Testing by YOLOv7 (based on Keelung River dataset) 64
4.2 Fixed and Controlled Variations of Datasets 72
4.2.1 Case 1: Fixed Variations of Datasets 72
4.2.2 Case 2: Control Variations of Datasets 79
4.3 Garbage Accumulation River Section in Keelung City 81
Chapter 5: Conclusion and Future Prospects 90
5.1 Conclusion 90
5.2 Improvements and Future Prospects 91
References 92


































List of Figures
Figure 1. Riverside pollution. 1
Figure 2. UAV system. 4
Figure 3. Proposed hexacopter. 5
Figure 4. PID parameters of the hexacopter [18]. 6
Figure 5. Flowchart of the study for the trash detection task. 7
Figure 6. Drone power supply. 8
Figure 7. The performance of the proposed hexacopter [21]. 8
Figure 8. YOLOs evolution [29] [30]. 9
Figure 9. Scaled-YOLOv4 results, inference speed on the X-axis, and COCO average precision on the Y-axis [5]. 10
Figure 10. YOLOv4 architectural [30]. 11
Figure 11. The results of YOLOv4 [31] are based on the Keelung River dataset. 12
Figure 12. YOLOv5 AP rate [32]. 13
Figure 13. YOLOv5 architecture [33]. 14
Figure 14. The results of YOLOv5 [32] are based on the Keelung River dataset. 16
Figure 15. Comparing various existing real-time object detectors. [12] 17
Figure 16. YOLOv7 architecture [36]. 18
Figure 17. Comparison between ELAN architecture and E-ELAN [12]. 19
Figure 18. The impact of scaling up depth on the series model [12]. 20
Figure 19. RepVGG training and inference architecture [12]. 23
Figure 20. ResNet has a residual architecture, so it is not suitable to use id to reorganize parameters [12]. 24
Figure 21. RepConv position [12]. 25
Figure 22. RDark block [12]. 26
Figure 23. The proposed label assigner is optimized by lead head prediction and the ground truth to get the labels of the training lead head and auxiliary head at the same time [12]. 27
Figure 24. Independent vs lead guided [12]. 28
Figure 25. a to f are YOLOv7, YOLOv7-d6, YOLOv7-e6, YOLOv7-e6e, YOLOv7-w6, and YOLOv7x, respectively, using the mAP results of the p5 parameter in the Keelung River dataset. 29
Figure 26. a to f are YOLOv7, YOLOv7-d6, YOLOv7-e6, YOLOv7-e6e, YOLOv7-w6, and YOLOv7x, respectively, using the mAP results of the p6 parameter in the Keelung River dataset. 30
Figure 27. a to f are YOLOv7, YOLOv7-d6, YOLOv7-e6, YOLOv7-e6e, YOLOv7-w6, YOLOv7x, and YOLOv7-tiny respectively, using the tiny parameter in the Keelung River dataset, and the size of the dataset is set at 640*640 mAP result. 31
Figure 28. a to c are YOLOv7-e6, YOLOv7-e6e, and YOLOv7-w6 respectively, using the tiny parameter in the Keelung River dataset, and the size of the dataset is set at 1280*1280 mAP result. 32
Figure 29. Keelung-River trash dataset. 34
Figure 30. Garbage and bottles in the Keelung-River trash dataset. 44
Figure 31. Negative samples in the Keelung-River trash dataset. 46
Figure 32. The real-world detection in the Keelung River (the YOLOv4 model is trained from the Keelung River dataset ). 55
Figure 33. The real-world detection in the Keelung River (the YOLOv5 model is trained from the Keelung River dataset ). 63
Figure 34. The real-world detection in the Keelung River (the YOLOv7 model is trained from the Keelung River dataset). 71
Figure 35. a to g are the results of seven architecture-labeled samples of YOLOv7. 73
Figure 36. a to g are the results of seven architecture-labeled samples of YOLOv7. 74
Figure 37. a to g are the results of seven architecture-labeled samples of YOLOv7. 75
Figure 38. a to g are the results of seven architecture-labeled samples of YOLOv7. 76
Figure 39. a to g are the results of seven architecture-labeled samples of YOLOv7. 77
Figure 40. a to g are the results of seven architecture-labeled samples of YOLOv7. 78
Figure 41. a to d is the detection value Img_size designed to be 640 to 5120. 81
Figure 42. From 2022 to 2023 in Xiding River, different structures have identified the garbage so as to facilitate the reference to the Environmental Protection Bureau for environmental maintenance. 82
Figure 43. From 2022 to 2023 in Tianliao River, different structures have identified the garbage so as to facilitate the reference to the Environmental Protection Bureau for environmental maintenance. 84
Figure 44. From 2022 to 2023 in Sijiao Pavilion, different structures have identified the garbage so as to facilitate the reference to the Environmental Protection Bureau for environmental maintenance. 85
Figure 45. From 2022 to 2023 in Shijian Bridge, different structures have identified the garbage so as to facilitate the reference to the Environmental Protection Bureau for environmental maintenance. 86
Figure 46. From 2022 to 2023 in Dahua Bridge, different structures have identified the garbage so as to facilitate the reference to the Environmental Protection Bureau for environmental maintenance. 87
Figure 47. From 2022 to 2023 in Bade Bridge, different structures have identified the garbage so as to facilitate the reference to the Environmental Protection Bureau for environmental maintenance. 88
Figure 48. Contact the Environmental Protection Bureau and the results after processing by the Environmental Protection Bureau. 89



























List of Tables
Table 1. The difference between YOLOv3 and YOLOv4 [4]. 11
Table 2. YOLOv5 v.s. YOLOv4 and YOLOv3, modified from [34]. 15
Table 3. Ablation study of compound model scaling [12]. 21
Table 4. Ablation study of planned RepConcatenation model [12]. 25
Table 5. Ablation study of planned RepResidual model [12]. 26
Table 6. Proposed auxiliary head ablation study [12]. 27
Table 7. The seven model architectures of YOLOv7 are based on the Keelung River dataset, and the real-time identification results are compiled. 33
Table 8. Comparison of garbage identification with different models. 79
Table 9. The parameters are adjusted in different places of p5, p6, and tiny parameter files [33]. 79
 

References
[1]T. Jia, Z. Kapelan, R. de Vries, P. Vriend, E. C. Peereboom, I. Okkerman, R. Taormina, “Deep learning for detecting macroplastic litter in water bodies: A review,” Proceedings of IEEE International Conference on Applied System Innovation, 2021.
[2]P. F. Proença and P. Simões, “TACO: Trash Annotations in Context for Litter Detection,” Mar.2020, [Online]. Available: http://arxiv.org/abs/2003.06975.
[3]“pedropro/TACO: Trash Annotations in Context Dataset Toolkit.” Access date: Feb. 2022, [Online]. Available: https://github.com/pedropro/TACO.
[4]“YOLOv4 Industrial Application Experience - Zhang Jiaming,” Access date: Apr. 2022, [Online]. Available: https://aiacademy.tw/yolo-v4-intro/.
[5]C.Y. Wang and H. Y. M. Liao, “Scaled-YOLOv4: Scaling Cross Stage Partial Network,” arXiv preprint arXiv:2011.08036, 2020.
[6]S. Merlino, M. Paterni, A. Berton, and L. Massetti, “Unmanned Aerial Vehicles for Debris Survey in Coastal Areas: Long-Term Monitoring Programme to Study Spatial and Temporal Accumulation of the Dynamics of Beached Marine Litter,” Remote Sensing 12, no. 8: 1260. https://doi.org/10.3390/rs12081260.
[7]E. S. Gabriela, H. Mirco, O. Natascha, and S. Gerald., “Efficiency of Aerial Drones for Macrolitter Monitoring on Baltic Sea Beaches,” Frontiers in Environmental Science, vol. 8, 2021, doi: 10.3389/fenvs.2020.560237.
[8]M. Tharani, A. W. Amin, M. Maaz, and, M.Taj, “Attention Neural Network for Trash Detection on Water Channels,” arXiv preprint arXiv:2007.04639, 2020.
[9]“What is YOLOv5? A Guide for Beginners.” Access date: Apr. 2022, [Online]. Available: https://blog.roboflow.com/yolov5-improvements-and-evaluation/.
[10]S. Xu, Z. Guo, Y. Liu, J. Fan, and X. Liu, “An Improved Lightweight YOLOv5 Model Based on AttentionMechanism for Face Mask Detection,” 11 Sep 2022 [cs.CV]. arXiv:2203.16506v3.
[11]“YOLOv7: The Most Powerful Object Detection Algorithm (2023 Guide),” Access date: Jul. 2022, [Online]. Available: https://viso.ai/deep-learning/yolov7-guide/.
[12]C.Y.Wang, A.Bochkovskiy, and H-Y.M.Liao, “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” 6.Jul.2022, [cs.CV]. arXiv:2207.02696v1.
[13]M. Tharani, A. W. Amin, M. Maaz, and M.Taj, “Attention Neural Network for Trash Detection on Water Channels,” arXiv preprint arXiv:2007.04639, 2020.

[14]C. L. Chang, K. Y. Chong, C. F. Chen, J. Y. Lin, and S. F. Kang “Research on pollution source analysis and potential impacts on water quality in the Keelung River, ” Journal of Taiwan Agricultural Engineering, vol. 66, no. 2, Jun. 2022, doi: 10.29974/JTAE.202206_68(2).0002.
[15]“DWM-222 4G LTE N150 USB Adapter.” Access date: Feb. 2022, [Online]. Available:https://eu.dlink.com/uk/en/-/media/consumer_products/dwm/dwm-222/datasheet/dwm_222_a2_datasheet_en_eu.pdf.
[16]“Holybro Pixhawk 6X” Access date: Jul. 2022, [Online]. Available: https://docs.px4.io/main/en/flight_controller/pixhawk6x.html.
[17]“ZED-F9P module.” Access date: Jul. 2022, [Online]. Available: https://www.u-blox.com/en/product/zed-f9p-module.
[18]“PID controller - Wikipedia.” Access date: Fed. 2022, [Online]. Available: https://en.wikipedia.org/wiki/PID_controller.
[19]“reComputer J2012 - Edge AI Device with Jetson Xavier NX 16 GB module, Aluminium case, pre-installed JetPack System” Access date: Feb. 2022, [Online]. Available: https://www.seeedstudio.com/Jetson-20-1-H2-p-5329.html.
[20]“SJCAM SJ5000X specifications.” Access date: Feb. 2022, [Online]. Available: https://www.manua.ls/sjcam/sj5000x/specifications.
[21]“Flying Evaluation,” Access date: Apr. 2022, [Online]. Available: https://flyeval.com/paper/.
[22]J. R. Terven, and D. M. Cordova-Esparaza, “A Comprehensive Review of YOLO: From YOLOv1 And Beyond,” Under Review in ACM Computing Surveys, arXiv:2304.00501v3 [cs.CV] 9 Jun. 2023, [Online]. Available: https://arxiv.org/pdf/2304.00501.
[23]R. Girshick, J. Donahue, T. Darrell, andJ. Malik, “Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation,” Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Sep. 2014, pp. 580–587.
[24]C. Y. Wang, H. Y. M. Liao, Y. H. Wu, P. Y. Chen, J. W. Hsieh, and I. H. Yeh, “CSPNet: A New Backbone that can Enhance Learning Capability of CNN,” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020.
[25]J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 2016-Decem, pp. 779–788, Jun.2015.
[26]J. Redmon and A. Farhadi, “YOLO9000: Better, Faster, Stronger,” Proceedings of 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, vol. 2017-Janua, pp. 6517–6525, 2017.
[27]J.Redmon and A.Farhadi, “YOLOv3: An Incremental Improvement,” Apr.2018, [Online]. Available: http://arxiv.org/abs/1804.02767.
[28] A. Bochkovskiy, C. Y. Wang, and H. Y. M. Liao, “YOLOv4: Optimal Speed and Accuracy of Object Detection,” arXiv preprint arXiv:2004.10934, 2020.
[29]“Evolution of YOLO family series models: from v1 to v8 (below),” Access date: Apr. 2022, [Online]. Available: https://segmentfault.com/a/1190000043338809.
[30]“YOLOv1-v4+PP-YOLO.” Access date: Apr. 2022, [Online]. Available: https://hackmd.io/@ZZ/yolov4-ppyolo.
[31]“pytorch-YOLOv4.” Access date: Apr. 2022, [Online]. Available: https://github.com/Tianxiaomo/pytorch-YOLOv4.
[32]“ultralytics/yolov5 v6.0_v6.0 - YOLOv5n 'Nano' models, Roboflow integration, TensorFlow export, OpenCV DNN support on GitHub.” Access date: May. 2022, [Online]. Available: https://newreleases.io/project/github/ultralytics/yolov5/release/v6.0.
[33]“A complete explanation of the core basic knowledge of Yolov5 in the Yolo series.” Access date: Apr. 2022, [Online]. Available: https://zhuanlan.zhihu.com/p/172121380.
[34]“Object Detection using YOLOv5 OpenCV DNN in C++ and Python.” Access date: Apr. 2022, [Online]. Available: https://learnopencv.com/object-detection-using-yolov5-and-opencv-dnn-in-c-and-python/.
[35]“ultralytics/yolov5.” Access date: Jun. 2022, [Online]. Available: https://github.com/ultralytics/yolov5.
[36]yolov7_custom_data_trainig.” Access date: Jul. 2022, [Online]. Available:https://github.com/akashAD98/yolov7_custom_data_trainig.
[37]M. Tan, R. Pang, and Q. V. Le, “EfficientDet: Scalable and Efficient Object Detection,” Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp. 10778–10787, Nov.2019.
[38]S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He, “Aggregated Residual Transformations for Deep Neural Networks,” Proceedings of the 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Nov. 2017, vol. 2017-January, pp. 5987–5995.