應用無人機及物件偵測於大園海灘的瓶裝海洋垃圾;Automatic detection of bottle marine debris on Dayuan beaches using unmanned aerial vehicles and machine learning techniques

NCUIR > College of Earth Sciences > Graduate Institute of Hydrological and Oceanic Sciences > Electronic Thesis & Dissertation > Item 987654321/85741

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/85741

Title:	應用無人機及物件偵測於大園海灘的瓶裝海洋垃圾;Automatic detection of bottle marine debris on Dayuan beaches using unmanned aerial vehicles and machine learning techniques
Authors:	陳氏玲慈;CHI, TRAN THI LINH
Contributors:	水文與海洋科學研究所
Keywords:	瓶裝海洋垃圾;無人機;背景去除;機器學習;YOLO v2;物件偵測;數據增強;bottle marine debris;UAV;data augmentation;machine learning;YOLO v2;object detection;background removal image
Date:	2021-09-28
Issue Date:	2021-12-07 11:20:31 (UTC+8)
Publisher:	國立中央大學
Abstract:	在現今社會中人們的環保意識日益增強，然而，瓶裝海洋垃圾 (BMD) 仍然是世界上備受重視的環境問題之一。傳統海灘垃圾研究中的監測方法因為人力資源的關係存在著許多局限性，因此本研究提出了一種利用無人機和物件辨識BMD的方法，在桃園市大園區的沙灘上進行了相關研究。首先，本研究設計了三個實驗區域進行圖像收集用於模型訓練，且為了確保此方法在長期研究上的可行性，另外收集了兩處真實區域(非實驗區域)之圖像用於驗證模型強健性。接著，使用無人機於不同高度收集圖像，其解析度為 0.12 至 1.54 厘米/像素；物件辨識系統則採用You Only Look Once version 2 (YOLO v2) ，其使用無人機收集之圖像進行訓練 BMD辨識模型；此外本研究應用背景移除之影像處理技術來移除圖像中的雜訊、於訓練過程中應用數據增強(Data augmentation)之技術增加訓練數據量以提升模型可信性，並採用聯合交集（IoU）來評估訓練效率。本研究發現在航測上使用 0.5 厘米/像素的解析度能得到最佳的結果，該解析度於實驗區域之準確率(precision)達到 0.94及召回率(recall rate)達到0.97 ，可得 0.95 的 F1-score；在真實區域上，檢測的平均準確率為 0.61，召回率為 0.86，F1-score為 0.72。本研究顯示，數據增強之應用在訓練過程中起著至關重要的作用，其結果IoU 超過 0.68；而背景移除技術則大量節省整個檢測時間，也因為移除了大量雜訊，減少了於真實區域中檢測錯誤的情況，證實數據增強及背景移除技術可以更準確、快速和客觀地識別海灘上的垃圾。;Humans’ awareness of the environment is increasing nowadays; however, bottle marine debris (BMD) remains one of the most pressing global issues. Fields surveys of marine debris based on manpower is less efficient; therefore, this study proposes an automatic detection method on BMD using unmanned aerial vehicles and machine learning techniques. The study sites are located on sandy beaches in Dayuan District, Taoyuan City. We first set three designed sites to create training datasets and test the detecting algorithm, and performances. Two real sites were then surveyed to evaluate our method in such a sandy complex beach that was intended to be used for long-term researches. The UAVs were operated at different fly heights to capture images with resolutions from 0.12 to 1.54 cm/pixel. The object detection algorithm You Only Look Once version 2 (YOLO v2) was trained to identify BMD and we added an image processing skill to remove image background noises. Data augmentation was used in training process to increase training data, and intersection over union (IoU) was adopted to evaluate the training efficiency. The results reveal that the skill of data augmentation helps IoU reaches over 0.68; and the skill of background removal has an advantage to reduce the processing time, as well as reducing noise resulting in much greater precision in real sites. From testing on both the designed and real sites with different image resolutions and processing skills, we found that approximately 0.5 cm/pixel could be the optimal resolution for aerial surveys on BMD. When operating the UAV with an image resolution of 0.5 cm/pixel, the performance indexes of mean precision, recall rate, and F1-score are respectively, 0.94, 0.97 and 0.95 at designed sites and are 0.61, 0.86, and 0.72 at real sites. Our work contributes to advances in beach debris surveys, optimizes the automatic detection on machine learning approach, especially with the role of data augmentation step in training data and background removing procedure.
Appears in Collections:	[Graduate Institute of Hydrological and Oceanic Sciences] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	122	View/Open

社群 sharing

Loading...