臺灣博碩士論文加值系統

近年來，影像辨識及物件偵測技術已逐漸成形，皆得益於卷積神經網路的技術與眾多研究人員的付出，現今此技術也被廣泛應用在各領域之上，影像辨識約可分為三種問題物件分類、物件定位及物件偵測，不同的影像辨識系統，其功能與辨識結果皆不盡相同。本文使用不同的深度學習網路YOLO以及ResNet，對水下物件四種類別（蛙人、潛艦、魚雷、水下探測設備）進行物件偵測辨識，以及對農地作物三種類別(葡萄、梨子、蓮霧)進行影像辨識。水下物件偵測方面，使用影像標註軟體框選水下物件，搭配YOLOv5與YOLOv7不同深度的神經網路架構，進行特徵訓練；農地作物辨識，則用有一定深度的神經網路架構ResNet50，透過隨機翻轉影像，抓取整張影像不同角度特徵來進行訓練。最終實驗結果，水下物件在YOLOv5以及YOLOv7所有類別物件偵測辨識中，各模型平均精確度及召回率都在75%以上；農地作物在ResNet50所有類別的影像辨識，各模型也都有平均85%以上的準確度。

In recent years, image recognition and object detection technology has gradually taken shape, thanks to the convolutional neural network technology and the efforts of many researchers, and now this technology is widely used in various fields. Image recognition can be divided into three kinds of problem object classification, object location and object detection. Different image recognition systems have different functions and identification results. In this paper, different deep learning networks YOLO and ResNet are used for object detection and identification of four categories of underwater objects (frogman, submarine, torpedo, underwater detection equipment), and image identification of three categories of agricultural crops (grape, pear, lotus fog). In terms of underwater object detection, underwater objects are selected by image annotation software frame, and feature training is conducted by matching neural network architecture of YOLOv5 and YOLOv7 at different depths. For crop identification of agricultural land, ResNet50, a neural network architecture with a certain depth, is used for training by randomly flipping images and capturing features from different angles of the whole image. The final experimental results showed that in the detection and identification of underwater objects in YOLOv5 and YOLOv7, the average accuracy and recall rate of each model were above 75%. For all categories of ResNet50 image identification of farm crops, the accuracy of each model is above 85% on average.

摘要
Abstract
誌謝
目錄
圖目錄
表目錄
第一章、緒論
1.1前言
1.2研究動機
1.3研究架構
第二章、文獻探討
2.1 YOLO
2.1.1 YOLO（You Only Look Once）
2.1.2 YOLOv2
2.1.3 YOLOv3
2.1.4 YOLOv4
2.1.5 YOLOv5
2.2 LabelImg
2.3水下影像問題
2.4 ResNet
第三章、研究方法及流程
3.1 研究方法
3.2 研究流程
3.2.1 篩分資料
3.2.2 標定圖像
3.2.3 圖像訓練及驗證
3.3 置信度與IoU值
3.4 過度訓練
第四章、實驗結果與討論
4.1 實驗環境
4.2 實驗資料
4.3 模型評估
4.4 實驗結果
第五章、結論
參考文獻

[1]Costanza Robert, "The ecological, economic, and social importance of the oceans." Ecological Economics, 1999.
[2]xahidbuffon, (Mar.10,2022),Awesome_Underwater_Datasets, Retrieved from: https://github.com/xahidbuffon/Awesome_Underwater_Datasets
[3]Robotics-Institute-HIST, (Jul.30,2020), Underwater Dangerous Target dataset, Retrieved from: https://github.com/Robotics-Institute-HIST/Dataset
[4]國網中心資料集平台,(2022年9月12日), 國立成功大學測量及空間資訊學系(AI CUP 2022農地作物現況調查影像辨識競賽-秋季賽),Retrieved from: https://scidm.nchc.org.tw/organization/geomatics-ncku-edu-tw
[5]Tzutalin, (Sep.23,2022), LabelImg, Retrieved from: https://github.com/tzutalin/labelImg
[6]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Deep residual learning for image recognition." CVPR, 2016.
[7]Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. "You only look once: Unified, real-time object detection." CVPR, 2016.
[8]Joseph Redmon and Ali Farhadi. "YOLO9000: better, faster, stronger." CVPR, 2017.
[9]Joseph Redmon and Ali Farhadi. "Yolov3: An incremental improvement." arXiv:1804.02767, 2018.
[10]Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. "Yolov4: Optimal speed and accuracy of object detection." arXiv:2004.10934, 2020.
[11]Ultralytics, (Jan.21,2023), YOLOv5, Retrieved from: https://github.com/ultralytics/yolov5
[12]Raimondo Schettini and Silvia Corchs. "Underwater Image Processing: State of the Art of Restoration and Image Enhancement Methods." EURASIP Journal on Advances in Signal Processing, 2010.
[13]周景春, 衛曉靖, and 史金余. "基於藍綠通道自適應色彩補償的水下圖像增強." Journal of Electronics & Information Technology, 2022.
[14]Wiki, (Oct.10,2022), Overfitting, Retrieved from: https://en.wikipedia.org/wiki/Overfitting
[15]Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. "YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors." arXiv:2207.02696, 2022.
[16]PyTorch, Retrieved from: https://pytorch.org/
[17]Jiuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting Liu, Xingxing Wang, Gang Wang, Jianfei Cai, and Tsuhan Chen. "Recent advances in convolutional neural networks." Pattern Recognition, 2018.
[18]Tommy Huang, (Jul.10,2018), 機器學習\統計方法:模型評估-驗證指標(validation index), Retrieved from: https://chih-sheng-huang821.medium.com/機器學習-統計方法-模型評估-驗證指標-b03825ff0814
[19]LeCun Yann, Bottou Léon, Bengio Yoshua, and Haffner Patrick. "Gradient-based learning applied to document recognition." IEEE Access, 1998.
[20]Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Communications of the ACM, 2017.
[21]Simonyan Karen and Andrew Zisserman. "Very Deep Convolutional Networks for Large-Scale Image Recognition." arXiv:1409.1556, 2015.
[22]Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. "Going deeper with convolutions." CVPR, 2015.
[23]Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. "Rich feature hierarchies for accurate object detection and semantic segmentation." CVPR, 2014.
[24]Girshick Ross. "Fast r-cnn." CVPR, 2015.
[25]Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. "Faster r-cnn: Towards real-time object detection with region proposal networks." NIPS 2015, 2015.
[26]Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. "Ssd: Single shot multibox detector." ECCV, 2016.
[27]Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. "Mask r-cnn." CVPR, 2017.