臺灣博碩士論文加值系統

臺灣作為海島國家，海運和漁業是其經濟發展的重要支柱。隨著全球化趨勢加速和國際貿易量的持續增長，臺灣港口面臨的營運壓力逐年上升，管理複雜度不斷提高。為了應對這一挑戰，港口管理迫切需要轉型升級，採用更智慧化、自動化的系統，以提升營運效率和降低對人工的依賴。進一步來說，開發和實施先進的智慧化和自動化管理系統成為提高臺灣海運和漁業競爭力、確保經濟安全的戰略性需求。這不僅能有效監控和管理船舶動態，同時也是優化港口營運、增強行業可持續性的關鍵舉措。
由於近年來人工智慧技術的迅速進步，特別是在影像處理和深度學習領域的創新，當前各種深度學習技術已經成為船舶辨識任務中不可或缺的工具，本研究透過使用PP-ShiTuV2，能夠以輕量化的資料集來進行訓練，且擁有很高的精度，能夠有效地進行切割影像和高效的模型，不僅能對多類別進行精確識別，也可以滿足對預測效率的極致追求，最終得到在CPU (Central Processing Unit)上可完成即時的影像識別的系統。總體而言，目前運用深度學習方法可大大的提升的漁船目標檢測的識別效能。
關鍵詞：深度學習、輕量級影像辨識、船舶辨識

As an island nation, Taiwan relies heavily on maritime transport and fisheries as crucial pillars of its economic development. With the acceleration of globalization and the continuous growth of international trade, the operational pressures on Taiwan's ports increase annually, and the complexity of management continues to rise. To address these challenges, there is an urgent need for port management to transform and upgrade by adopting more intelligent and automated systems to enhance operational efficiency and reduce reliance on manual labor. Moreover, the development and implementation of advanced intelligent and automated management systems are becoming strategic necessities to enhance the competitiveness of Taiwan's maritime and fisheries sectors and ensure economic security. These measures not only effectively monitor and manage vessel dynamics but also optimize port operations and enhance the sustainability of the industry.
Due to the rapid advancement in artificial intelligence technology, particularly innovations in image processing and machine learning, various deep learning techniques have become indispensable tools for ship recognition tasks. In this study, by employing PP-ShiTuV2, training can be conducted using lightweight datasets with high accuracy. The system effectively processes and models segmented images, capable of precisely identifying multiple categories and meeting the ultimate demands for prediction efficiency. Ultimately, obtain a system capable of performing real-time image recognition on a CPU(Central Processing Unit). Currently, the application of deep learning methods can significantly enhance the recognition performance of fishing vessel target detection.
Keywords: Deep Learning, Lightweight Image Recognition, Vessel Identification

目錄
摘要i
Abstract ii
誌謝 iii
表目錄 vi
圖目錄 vii
第一章 緒論 1
1.1研究背景與動機 1
1.2研究目的 2
1.3論文架構 2
第二章 文獻探討 3
2.1傳統船舶辨識 3
2.2 現代船舶辨識應用 5
2.2.1 YOLOv3船舶辨識 6
2.2.2 YOLOv3船舶字元辨識 7
2.2.3 YOLOv4船舶辨識 8
2.3 PP-ShiTuV2架構 9
2.3.1 Faster R-CNN 11
2.3.2 EfficientDet 12
2.3.3 PP-YOLOv2 13
2.4 特徵提取架構 14
2.4.1 Swin-Transformer 15
2.4.2 EfficientNet 16
2.4.3 PP-LCNetV2 17
第三章 研究方法 18
3.1 資料集 19
3.1.1 資料集蒐集 19
3.1.2 資料集樣本 20
3.1.3 資料預處理 21
3.2 PP-ShiTuV2系統 22
3.2.1 主體檢測 22
3.2.2 特徵提取 25
3.2.3 向量搜索 27
第四章 研究結果與分析 30
4.1 拍攝設備 30
4.2 資料集 32
4.3 環境配置 32
4.4 模型訓練模型訓練 33
4.5 評估指標評估指標 33
4.6 驗證結果比較驗證結果比較 36
第五章 結論與未來展望結論與未來展望 41
5.1結論結論 41
5.2 未來展望未來展望 41
參考文獻 42

[1]Wei, S. (2021). PP-ShiTu: A Practical Lightweight Image Recognition System. arXiv preprint arXiv:2111.00775.
[2]Ren, S., He, K., Girshick, R., & Sun, J. (2017). Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6), 1137-1149.
[3]He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN. arXiv preprint arXiv:1703.06870.
[4]Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 580-587).
[5]Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2015). You Only Look Once: Unified, Real-Time Object Detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 779-788).
[6]Redmon, J., & Farhadi, A. (2017). YOLO9000: Better, Faster, Stronger. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 6517-6525).
[7]Redmon, J., & Farhadi, A. (2018). YOLOv3: An Incremental Improvement. arXiv preprint arXiv:1804.02767.
[8]Bochkovskiy, A., Wang, C.Y., & Liao, H.Y. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv preprint arXiv:2004.10934.
[9]Tan, M., Pang, R., & Le, Q.V. (2019). EfficientDet: Scalable and Efficient Object Detection. arXiv preprint arXiv:1911.09070.
[10]劉致圻. (2019). 應用深度學習於船舶影像分類研究 [Master's thesis, National Taiwan Ocean University].
[11]林修賢. (2021). YOLO深度學習網路應用於漁船編號之辨識 [Master's thesis, National Taiwan Ocean University].
[12]Chen, R.-H. (2021). Research on deep neural networks based fishing vessels identification system [Master's thesis, National Chin-Yi University of Technology].
[13]Huang, X. (2021). PP-YOLOv2: A Practical Object Detector. arXiv preprint arXiv:2104.10419.
[14]Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., & Zhang, L. (2021). Swin Transformer: Hierarchical vision transformer using shifted windows. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, 13131-13140.
[15]Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking model scaling for convolutional neural networks. Proceedings of the 36th International Conference on Machine Learning (ICML), 2019, 6105-6114.
[16]Cui, C., Gao, T., Wei, S., et al. (2021). PP-LCNet: A Lightweight CPU Convolutional Neural Network. arXiv preprint arXiv:2109.15099.
[17]He, K., Zhang, X., Ren, S., & Sun, J. (2015). Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385. Retrieved from https://arxiv.org/abs/1512.03385
[18]Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017). Focal loss for dense object detection. arXiv preprint arXiv:1708.02002. Retrieved from https://arxiv.org/abs/1708.02002
[19]Facebook Engineering. (2017, March 29). Faiss: A library for efficient similarity search. Retrieved from https://engineering.fb.com/2017/03/29/data-infrastructure/faiss-a-library-for-efficient-similarity-search/
[20]Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L.-C. (2018). MobileNetV2: Inverted residuals and linear bottlenecks. arXiv preprint arXiv:1801.04381. Retrieved from https://arxiv.org/abs/1801.04381
[21]Ma, N., Zhang, X., Zheng, H. T., & Sun, J. (2018). ShuffleNet V2: Practical guidelines for efficient CNN architecture design. Proceedings of the European Conference on Computer Vision (ECCV), 2018, 116-131.
[22]Howard, A., Sandler, M., Chu, G., Chen, L.-C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., Le, Q. V., & Adam, H. (2019). Searching for MobileNetV3. *arXiv preprint arXiv:1905.02244*. Retrieved from https://arxiv.org/abs/1905.02244
[23]DC Fever. (n.d.). Sony Alpha 7R IV specifications. Retrieved July 2, 2024, from https://www.dcfever.com/cameras/fullspecification.php?id=2306
[24]Samsung. (n.d.). Galaxy S22 Ultra 5G (256GB). Retrieved July 2, 2024, from https://www.samsung.com/tw/smartphones/galaxy-s/galaxy-s22-ultra-burgundy-256gb-sm-s9080drgbri/
[25]Redmon, J., Farhadi, A. (2018). YOLOv3: An Incremental Improvement. arXiv preprint arXiv:1804.02767.
[26]蘇柏廷. (2023). 應用於肺腺癌病理切片之影像分割 [National Kaohsiung University of Science and Technology].