由於海洋佔地球面積約71%，人類自古以來不斷地探索這未知領域。然而水下探測技術逐漸成熟，帶動著各類應用技術的開發與探討。隨著能源轉型議題逐漸受到各界重視，我國政府於近年積極於苗栗、彰化、雲林一帶建設離岸風電場。然而，這類建設與許多水下環境保護議題有密切關聯，例如：中華白海豚以及野生動物的重要棲地均緊鄰離岸風電場址。評估離岸風電場址是否會破壞白海豚棲息地，成為關鍵研究議題之一。為了有效評估場址水文與水下生態環境，側掃聲納影像則是一項重要的工具之一。但在進行側掃聲納作業時，須仰賴專業觀測員長期注視側掃聲納影像，這不但容易造成眼睛疲勞進而影響水下目標物識別的精確度，且在調查的過程可能還會受到水下環境干擾，使得在作業時無法準確偵測拖魚與海床的作業高度。因此，本論文利用深度學習結合影像分析技術提出水下物件辨識的方法。但因在處理側掃聲納影像內容時，需先進行海床邊線偵測以便找出物件辨識的有效範圍。所以本論文將先詳細說明海床邊線偵測方法，在說明如何利用YOLOv3技術識別物件協助水下場址調查，輔佐觀測員即時的環境調查與紀錄。此法不但可減輕人力成本與觀測員疲憊度，降低設備遺失之風險，更能提高水下目標物識別的準確性。實驗結果表示，所提方法不但可有效偵測海床邊線，還可透過YOLOv3物件識別技術辨識魚群、石塊、…等物件，有助於水下環境調查。其中海床邊線偵測其左舷與右舷平均連續率提升87.8%及94.8%之正確性，而在對稱率也達到80.7%之正確性。而在YOLOv3的部分透過影像前處理以及類別標籤之定義以及數據增強的方式，其平均精確率也有達到72.3%，且在識別魚群的部分也有相當高的信心值。本論文結合以上兩種方法，平均計算一張畫框(2558*1278)只需0.023秒，證實此方法能即時輔佐觀測員進行觀測，且有良好的影像處理效果。因此，本論文所提出的海床邊線偵測與YOLOv3物件識別技術，可有效運用於離岸風電潛力場址的評估，以及未來離岸風機的水下環境監測。

Since ocean occupies about 71% of the earth's surface, humans have been exploring this unknown territory since ancient time. However, underwater detection technology has gradually matured, driving the development and discussion of various application technologies. As the issue of green energy has gradually received attention from people, Taiwan government has been actively building offshore wind farms in Miaoli, Zhanghua, and Yunlin in recent years. However, this type of construction is closely related to many underwater environmental protection issues. For example, Sousa chinensis and important habitats of wild animals are all close to coasts. Evaluating whether offshore wind farm sites will damage the habitat of Sousa chinensis has become one of the key research topics.
In order to effectively assess the site's underwater culture and underwater ecological environment, side scan sonar images are one of the important tools. However, when performing side scan sonar operations, it is necessary to rely on professional observers to watch the side scan sonar images for a long time. This not only easily causes eye fatigue and affects the accuracy of underwater target recognition, but also may be affected by water during the investigation process. Under environmental interference, it is impossible to accurately detect the height of the tow fish and the seabed during operation. Therefore, this paper uses deep learning combined with image analysis technology to propose a method for underwater object recognition. However, when processing side scan sonar images, seabed edge detection is required to find out the effective range of object recognition. Therefore, this paper will first explain the seabed edge detection method in detail, and then explain how to use YOLOv3 technology to detect objects to assist in underwater site surveys and assist observers in real-time environmental surveys and records. This method can not only reduce labor costs and observer fatigue, reduce the risk of equipment loss, but also improve the accuracy of underwater target recognition. The experimental results show that the proposed method can not only effectively detect the edge of the seabed, but also detect shoal, rocks, ... and other objects through YOLOv3 object recognition technology, which is helpful for underwater environment investigation. Among them, the average continuous rate of the port side and starboard side of the seabed detection is increased by 87.8% and 94.8% accuracy, and the symmetry rate also reaches 80.7% accuracy. In the part of YOLOv3, the average accuracy of the image pre-processing, the definition of category labels and the method of data enhancement has reached 72.3%, and there is also a very high confidence value in the part of detecting fish. This paper combines the above two methods, and it takes an average of 0.023 seconds to calculate a frame (2558*1278), which proves that this method can assist the observer to make observations immediately and has good effects after certain image-processing. Therefore, the seabed edge detection and YOLOv3 object identification technology proposed in this paper can be effectively applied to the assessment of offshore wind power potential sites and the underwater environment monitoring of offshore wind turbines in the future.

中文學位中文學位論文考試審定書 ii
摘要 iii
Abstract v
致謝 vii
第一章 緒論 1
1.1 前言 1
1.2 研究動機與目的 5
1.3 論文架構 7
第二章 文獻探討與技術分析 9
2.1 國內外離岸風電水下場址環境評估議題 9
2.2 側掃聲納系統原理與限制 12
2.2.1 系統原理介紹 12
2.2.2 聲學陰影 18
2.2.3 聲納訊號干擾 19
2.2.4 海床邊線之對稱與連續率 25
2.3 側掃聲納影像處理 27
2.3.1 聲學與影像訊號轉換 27
2.3.2 影像雜訊濾除 28
2.3.3 海床邊線偵測技術探討 29
2.4 基於深度學習之水下目標物識別技術 37
2.5 小結 40
第三章 側掃聲納影像分析技術 41
3.1 研究架構 41
3.2 側掃聲納影像之前處理與海床邊線偵測技術 42
3.2.1 影像銳化 44
3.2.2 影像模糊 47
3.2.3 OTSU二值化方法 49
3.2.4 海床邊線搜尋 52
3.3 水下目標物識別之深度學習模型 56
3.3.1 數據增強 57
3.3.2 側掃聲納影像錯誤標記濾除 59
3.3.3 模型訓練架構 62
3.4 小結 64
第四章 實驗結果與分析 65
4.1 實驗目的 65
4.2 實驗方法 65
4.2.1 實驗環境規格 65
4.2.2 模擬架構 66
4.2.3 水下目標物位置推算 67
4.3 海床邊線偵測演算法結果 68
4.4 深度學習(YOLOv3)實驗結果 72
4.4.1 不同類別標籤偵測比較 72
4.4.2 有無數據前處理差異比較 79
4.5 實驗數據分析 85
第五章 結論與展望 88
參考文獻 90
附件一　論文審查意見 96
附件二　論文比對結果 98
附件三　個人經歷與著作 101

[1]經濟部能源局,“107年我國燃料燃燒二氧化碳排放統計與分析”, [Online]. Available：https://reurl.cc/GV0gXD. [Accessed 2019].
[2]許向華、呂惠雯, “台灣可再生能源產業現況與未來展望-以風力發電和太陽光電為例＂, 2019 Win the PRIDE：用指標說故事.
[3]風險社會與政策研究中心, “2019臺灣能源情勢回顧”, [Online]. Available：https://reurl.cc/vDvk7e. [Accessed 2019].
[4]工業技術研究院, “風力發電服務窗口”,　 [Online]. Available：https://www.twtpo.org.tw/. [Accessed 2020].
[5]環境資訊中心, “離岸風機示範機組「福海」 選址衝擊生態 遭打入二階環評”, [Online]. Available：https://e-info.org.tw/node/206229. [Accessed 2019].
[6]環境資訊中心, “風機的水下世界 從風場開始思考海洋 學者：需長期監測與永續管理”, [Online]. Available：https://e-info.org.tw/node/220899. [Accessed 2020].
[7]W. J. Richardson, C. R. Greene, C. I. Mame, and D. H. Thomson, “Marine Mammals and Noise”. Academic Press: San Diego, CA, 1995.
[8]J. Carstensen, O. D. Henriksen, J. Teilmann, “Impacts of offshore windfarm construction on harbour porpoises: acoustic monitoring of echolocation activity using porpoise detectors (T-PODs)”, Marine Ecology. Prog. Ser, 321, pp.295–308, September, 2006
[9]J. Tougaard, P. T. Madsen, M. Wahlberg, “Underwater noise from construction and operation of offshore wind farms”, Bioacoustics, Vol.17, pp. 143–146, April, 2008
[10]H. Bailey, B. Senior, D. Simmons, J. Rusin, G. Picken, P. M. Thompson, “Assessing underwater noise levels during pile-driving at an offshore windfarm and its potential effects on marine mammals”, Marine Pollution Bulletin, Vol.60, pp.888–897, June, 2010
[11]M. J. Brandt, A. Diederichs, K. Betke and G. Nehls, “Responses of harbour porpoises to pile driving at the Horns reef II offshore wind farm in the Danish North Sea”, Marine Ecology Progress. Ser.421, pp.205–216, January, 2011
[12]F. H. Jensen, L. Bejder, M. Wahlberg, N A Soto, M Johnson, P T Madsen, “Vessel noise effects on delphinid communication”, Marine Ecology Progress. Ser. 395, pp.161–75, January, 2009
[13]M. Scheidat, J. Tougaard, S. Brasseur, J. Carstensen, T. V. P. Petel, J. Teilmann, P. Reijnders, “Harbour porpoises (Phocoena phocoena) and wind farms: a case study in the Dutch North Sea”, Journal of Environmental Research Letters. 6 025102 , June, 2011
[14]I. Spiga, J. Fox, R. Benson, “Potential effects of long-term exposure to boat noise on the growth, survival and nutrient retention in juvenile fish”, eds. The Effects of Noise on Aquatic Life. Springer, New York, pp. 255–257, 2012
[15]M. Wahlberg, H. Westerberg, “Hearing in fish and their reactions to sounds from offshore wind farms”, Marine Ecology Progress. Ser, 288, pp.295–309, March, 2005
[16]A. H. Auld, J. R. Schubel , “Effects of suspended sediment on fish eggs and larvae: a laboratory assessment Estuar”, Estuarine and Coastal Marine Science, vol.6, pp.153–164, February, 1978
[17]R. G. Lake, S. G. Hinch, “Acute effects of suspended sediment angularity on juvenile coho salmon (Oncorhynchus kisutch)”, Canadian Journal of Fisheries and Aquatic Sciences. Aquat. Sci, vol.56, pp.862–867, April, 1999
[18]G. J. Partridge, R. J. Michael, “Direct and indirect effects of simulated calcareous dredge material on eggs and larvae of pink snapper Pagrus auratus”, Journal of Fish Biology, vol.77, pp.227–40, July, 2010
[19]J. Teilmann, J. Carstensen, “Negative long term effects on harbour porpoises from a large scale offshore wind farm in the Baltic - evidence of slow recovery”, Journal of Environmental Research Letters.7: 045101, pp.1-11, December, 2012
[20]林子皓, 林思瑩, 余欣怡, 周蓮香, 離岸風場開發噪音對沿岸鯨豚類動物之衝擊, 風電月刊
[21]Bay Dynamics, “Starfish 990F Advanced Towed Side scan Sonar”, [Online]. Available : https://reurl.cc/exyXLj. [Accessed 2019].
[22]C. Mazel, “Side Scan Sonar Training Manual”, Klein Associates, Inc., pp.144, 1985
[23]Wikipedia, “side scan sonar” , [Online]. Available： https://reurl.cc/D9Lza5. [Accessed 2019].
[24]B. W. Flemming, “Side-Scan Sonar: A Practical Guide”, The international hydrographic review, Rev.53, pp.65-92, 1976
[25]E. Dura, S. Rakheja, L. Honghai and N. Kolev, “Image Processing Techniques For the Detection and Classification of Man Made Objects in Side-Scan Sonar Images”, Sonar Systems Nikolai Kolev, IntechOpen, DOI：10.5772/21920, pp.121-138, September, 2011
[26]S. Karabchevsky, “Real-Time Underwater Obstacle Detection Using Forward Looking Sonar and FPGA”, Ben-Gurion University of the Negev, DOI：10.1109,　September, 2011
[27]Wikipedia, “Feature scaling” , [Online]. Available： https://reurl.cc/z843mQ. [Accessed 2019].
[28]P. Sahu, N. Gupta and N. Sharma, “A survey on underwater image enhancement technique”, International Journal of Computer Applications, Vol.87, No.13, pp. 19-23, February, 2014
[29]G. Padmavathi, P. Subashini, M. M. Kumar and S. K. Thakur, “Comparison of filter used for underwater image preprocessing”, IJCSNS International Journal of Computer Science and Network Security, Vol.10, No.1, pp. 58-65, January, 2010
[30]S. Kumari and R. Vijay, “Effect of symlet filter order an denoising of still images”, Advanced Computing: An International Journal, Vol.3, No.1, pp. 137 – 143, February, 2012
[31]C. J. Prabhakar, P. U. Kumar, “An image based technique for enhancement of underwater image”, International Journal of Machine Intelligence, Vol.3, pp. 217–224, December, 2012
[32]A. A. Yassin, R. M. Ghadbanand, S. F. Saleh, “Using discrete wavelet transformation to enhance underwater image”, IJCSI International Journal of Computer Science Issues, Vol.10, pp. 220-228, September , 2013
[33]R. Kumudham, Dhanalakshmi, A Swaminathan, Geetha, V Rajendran, “Comparison of the performance metrics of median filter and wavelet filter when applied on sonar images for denoising”, 2016 International Conference on Computation of Power Energy Information and Commuincation (ICCPEIC), pp. 288-290, April, 2016.
[34]Z. Jianhu, W. Xiao, Z. Hongmei, W. Aixue, “A Comprehensive Bottom-Tracking Method for Side Scan Sonar Image Influenced by Complicated Measuring Environment” , in Proc. IEEE Journal of Ocenic Engineeing, Vol.42, pp. 619-631, November, 2016
[35]M. P. Hayes, T. Y. Ho, “Height Estimation of a Sonar Towfish from Side Scan Imagery”, Image and Vision Computing New Zealand, Hamilton, New Zealand, pp.102- 107, November, 2000.
[36]蘇加富, 洪盟峰, 謝欽旭, 施建州, “自適應側掃聲納影像盲區偵測演算法”, 2018 International Teaching Assistant Orientation, Penghu, Taiwan (R.O.C), pp1-6, May, 2018
[37]C. C. Shih, M. F. Horng, Y. R. Tseng, C. F. Su, C. Y. Chen, “An Adaptive Bottom Tracking Algorithm for Side-Scan Sonar Seabed Mapping”, in Proc. IEEE. Conf. 2019 IEEE Underwater Technology (UT), Kaohsiung, Taiwan, pp1-7, April, 2019
[38]F. Yu , Y. Zhu, Q. Wang, K. Li, M. Wu, Li, T. Yan, B. He, “Segmentation of Side Scan Sonar images on AUV”, in Proc. IEEE. Conf. Underwater Technology (UT), Kaohsiung, Taiwan, pp.1-4, April. 16-19, 2019
[39]Y. Song, Y. Zhu, G. Li, C. Feng, B. He, “Side Scan Sonar Segmentation Using Deep Convolutional Neural Network”, in Proc. IEEE. Conf. OCEANS 2017, Anchorage, AK, USA, pp.1-4, September. 18-21, 2017
[40]J. Rhinelander, “Feature Extraction and Target Classification of Side-Scan Sonar Images”, in Proc. IEEE. Conf. Symposium Series on Computational Intelligence (SSCI), Athens, Greece, pp. 1-6, December. 6-9, 2016
[41]X. W. Luo, X. M. Qin, Z. Y. Wu, F. L. Yang, M. W. Wang, J. H. Shang, “Sediment Classification of Small-Size Seabed Acoustic Images Using Convolutional Neural Networks”, in Proc. IEEE Access, Vol.7, pp. 98331-98339, July. 8, 2019
[42]D. Einsidler, M. Dhanak, P. P. Beaujean, “A Deep Learning Approach to Target Recognition in Side-Scan Sonar Imagery”, in Proc. IEEE. Conf . OCEANS 2018, Charleston, SC, USA, pp. 1-4, October. 22-25, 2018
[43]N. Otsu, “A Threshold Selection Method from Gray-Level Histograms”, IEEE Transactions on Systems, Man, and Cybernetics, 9, No. 1, pp.62-66, January, 1979
[44]J. Redmon, A. Farhadi, "YOLOv3: An incremental improvement" in arXiv:1804.02767, Apr. 2018, [online] Available: https://arxiv.org/abs/1804.02767. [Accessed 2019].