學術產出-Theses
Article View/Open
Publication Export
-
題名 基於多輸出標籤概念之階層式深度神經網路架構
Hierarchical Deep Neural Network Architecture based on Multi-Output Concept作者 索立桐
So, Li-Tung貢獻者 廖文宏
Liao, Wen-Hung
索立桐
So, Li-Tung關鍵詞 深度學習
多輸出分類器
階層式分類
階層一致性
預測風險
Deep learning
multi-output classifier
hierarchical classification
hierarchical consistency
prediction risk日期 2023 上傳時間 2-Jan-2024 15:39:30 (UTC+8) 摘要 階層式深度神經網路是一種能強化現有直觀平坦分類的深度學習框架，有效確保模型在學習的路線上路徑正確（由粗分類至細分類），並能在標註資料有限的情況下，針對同一目標拆分階層後，同時進行各面向的解析，使得預測出的結果可信度更高。 在一般的分類任務中，資料直接在某一種共通的條件下被定義出這筆資料的類別並賦予標籤，並沒有所謂的階層概念，造成容易區分的資料類型和難以區分的資料類型混雜在一起，導致後續的分類訓練上僅依靠訓練模型本身的擬合能力，如對資料進行階層區分後訓練，應能提升訓練的品質。 本論文旨在探討於圖像分類任務中使用階層式深度學習方式，展示其能帶來的各種訓練效益，我們提出使用多輸出概念，將多階層的分類問題轉化成為一般平坦分類器可容易修改處理的形式，並加上階層一致性、預測風險等限制，期使直接而有效地套用至現有的各類分類訓練模型，調整並觀察各階層回傳的損失梯度權重，並與一般的分類方法相互比較，最後研究結果指出，使用階層式深度學習方式不僅最終的準確率不低於原始直接分類的方式，而且在我們所定義的各類指標上都得出較一般方式更佳的結果。
Hierarchical Deep Neural Networks are a type of deep learning framework that enhances existing intuitive flat classifications. They effectively ensure that the model follows the correct learning path, i.e., coarse-to-fine classification. Even with limited labeled data, they analyze various aspects simultaneously by splitting the training sets into sub-classes according to a predefined hierarchy. This approach increases the credibility of the predicted results. In typical classification tasks, data is directly assigned to a category and labeled based on a common set of conditions, without the concept of hierarchy. This leads to a mixture of easily distinguishable and difficult-to-distinguish data types, relying solely on the fitting capability of the training model in subsequent classification training. Therefore, training data with hierarchical divisions should improve the quality of training. This thesis aims to explore the use of hierarchical deep learning in image classification tasks and demonstrate the various training benefits it brings. We propose the utilization of the multi-label concept to convert multi-level classification problems into a form that can be easily modified and processed by general flat classifiers. We also incorporate constraints such as hierarchical consistency and prediction risk to enable direct and effective application to various existing classification training models. During the process, we adjust and observe the loss gradients weights returned by each hierarchy level, and compare them with conventional classification methods. The research results indicate that using the hierarchical deep learning approach not only achieves accuracy comparable to or higher than the original direct classification method but also yields better results in various performance indicators.參考文獻 [1] 朱家宏. 階層式深度神經網路及其應用, pages 18-20, 2023. [2] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016. [3] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1-9), 2015. [4] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv 2017, arXiv: 1704.04861.2017. [5] Dmitry Retinskiy. (2020). Multi-Label Image Classification with PyTorch. Retrieved from https://learnopencv.com/multi-label-image-classification-with-pytorch(Oct.13,2023) [6] CNN for deep learning | Convolutional neural networks. Retrieved from https://datapeaker.com/en/big--data/cnn-for-deep-learning-convolutional-neural-networks(Oct.13,2023) [7] LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. [8] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6), 84-90. [9] K. Simonyan and A. Zisserman, Very deep convolutional networks for large-scale image recognition, arXiv preprint arXiv: 1409.1556., 2014. [10] Yan, Z., Zhang, H., Piramuthu, R., Jagadeesh, V., DeCoste, D., Di, W., & Yu, Y.. HD-CNN: hierarchical deep convolutional neural networks for large scale visual recognition. In Proceedings of the IEEE international conference on computer vision (pp. 2740-2748). 2015. [11] Xinqi Zhu, Michael Bain. B-CNN: Branch Convolutional Neural Network for Hierarchical Classification. arXiv 2017, arXiv:1709.09890.2017. [12] Salma Taoufiq , Balázs Nagy , Csaba Benedek.HierarchyNet: Hierarchical CNN-Based Urban Building Classification. Remote Sensing,Volume 12 ,Issue 22 .2020. [13] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Computer Science Department, University of Toronto, Tech. Rep, 2009. [14] D. Arthur and S. Vassilvitskii. K-means++: The advantages of careful seeding. In SODA, pages 1027–1035, 2007. 描述 碩士
國立政治大學
資訊科學系碩士在職專班
107971024資料來源 http://thesis.lib.nccu.edu.tw/record/#G0107971024 資料類型 thesis dc.contributor.advisor 廖文宏 zh_TW dc.contributor.advisor Liao, Wen-Hung en_US dc.contributor.author (Authors) 索立桐 zh_TW dc.contributor.author (Authors) So, Li-Tung en_US dc.creator (作者) 索立桐 zh_TW dc.creator (作者) So, Li-Tung en_US dc.date (日期) 2023 en_US dc.date.accessioned 2-Jan-2024 15:39:30 (UTC+8) - dc.date.available 2-Jan-2024 15:39:30 (UTC+8) - dc.date.issued (上傳時間) 2-Jan-2024 15:39:30 (UTC+8) - dc.identifier (Other Identifiers) G0107971024 en_US dc.identifier.uri (URI) https://nccur.lib.nccu.edu.tw/handle/140.119/149072 - dc.description (描述) 碩士 zh_TW dc.description (描述) 國立政治大學 zh_TW dc.description (描述) 資訊科學系碩士在職專班 zh_TW dc.description (描述) 107971024 zh_TW dc.description.abstract (摘要) 階層式深度神經網路是一種能強化現有直觀平坦分類的深度學習框架，有效確保模型在學習的路線上路徑正確（由粗分類至細分類），並能在標註資料有限的情況下，針對同一目標拆分階層後，同時進行各面向的解析，使得預測出的結果可信度更高。 在一般的分類任務中，資料直接在某一種共通的條件下被定義出這筆資料的類別並賦予標籤，並沒有所謂的階層概念，造成容易區分的資料類型和難以區分的資料類型混雜在一起，導致後續的分類訓練上僅依靠訓練模型本身的擬合能力，如對資料進行階層區分後訓練，應能提升訓練的品質。 本論文旨在探討於圖像分類任務中使用階層式深度學習方式，展示其能帶來的各種訓練效益，我們提出使用多輸出概念，將多階層的分類問題轉化成為一般平坦分類器可容易修改處理的形式，並加上階層一致性、預測風險等限制，期使直接而有效地套用至現有的各類分類訓練模型，調整並觀察各階層回傳的損失梯度權重，並與一般的分類方法相互比較，最後研究結果指出，使用階層式深度學習方式不僅最終的準確率不低於原始直接分類的方式，而且在我們所定義的各類指標上都得出較一般方式更佳的結果。 zh_TW dc.description.abstract (摘要) Hierarchical Deep Neural Networks are a type of deep learning framework that enhances existing intuitive flat classifications. They effectively ensure that the model follows the correct learning path, i.e., coarse-to-fine classification. Even with limited labeled data, they analyze various aspects simultaneously by splitting the training sets into sub-classes according to a predefined hierarchy. This approach increases the credibility of the predicted results. In typical classification tasks, data is directly assigned to a category and labeled based on a common set of conditions, without the concept of hierarchy. This leads to a mixture of easily distinguishable and difficult-to-distinguish data types, relying solely on the fitting capability of the training model in subsequent classification training. Therefore, training data with hierarchical divisions should improve the quality of training. This thesis aims to explore the use of hierarchical deep learning in image classification tasks and demonstrate the various training benefits it brings. We propose the utilization of the multi-label concept to convert multi-level classification problems into a form that can be easily modified and processed by general flat classifiers. We also incorporate constraints such as hierarchical consistency and prediction risk to enable direct and effective application to various existing classification training models. During the process, we adjust and observe the loss gradients weights returned by each hierarchy level, and compare them with conventional classification methods. The research results indicate that using the hierarchical deep learning approach not only achieves accuracy comparable to or higher than the original direct classification method but also yields better results in various performance indicators. en_US dc.description.tableofcontents 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 論文貢獻 2 1.4 論文架構 3 第二章 相關研究與技術背景 4 2.1 深度學習之圖像分類 5 2.2 階層式神經網路發展背景與技術 12 2.2.1 HD-CNN 架構 13 2.2.2 B-CNN架構 14 2.2.3 HierarchyNet架構 15 2.3 多輸出標籤概念 16 2.4 分群演算法 18 2.5 階層式分類評估指標 19 第三章 研究方法 21 3.1 訓練資料的前處理 21 3.2 模型及訓練方法 22 3.2.1 多輸出分類模型 22 3.2.2 階層式訓練方法 23 3.2.3 評估指標 26 3.3 設計概念 26 3.4 預期目標 29 第四章 實驗過程與結果分析 30 4.1 資料集 30 4.2 訓練資料的前處理 31 4.3 模型及訓練方法 33 4.3.1 基準模型 34 4.3.2 基準模型結合一般訓練方式 34 4.3.3 多輸出分類模型 35 4.3.4 多輸出分類模型結合階層式訓練方式 36 4.3.5 評估指標分析 51 4.4 達成的目標 57 4.4.1 可解釋性分析 58 4.4.2 可利用的特性 60 第五章 結論與未來研究方向 61 5.1 結論 61 5.2 未來研究方向 61 參考文獻 63 zh_TW dc.format.extent 3760158 bytes - dc.format.mimetype application/pdf - dc.source.uri (資料來源) http://thesis.lib.nccu.edu.tw/record/#G0107971024 en_US dc.subject (關鍵詞) 深度學習 zh_TW dc.subject (關鍵詞) 多輸出分類器 zh_TW dc.subject (關鍵詞) 階層式分類 zh_TW dc.subject (關鍵詞) 階層一致性 zh_TW dc.subject (關鍵詞) 預測風險 zh_TW dc.subject (關鍵詞) Deep learning en_US dc.subject (關鍵詞) multi-output classifier en_US dc.subject (關鍵詞) hierarchical classification en_US dc.subject (關鍵詞) hierarchical consistency en_US dc.subject (關鍵詞) prediction risk en_US dc.title (題名) 基於多輸出標籤概念之階層式深度神經網路架構 zh_TW dc.title (題名) Hierarchical Deep Neural Network Architecture based on Multi-Output Concept en_US dc.type (資料類型) thesis en_US dc.relation.reference (參考文獻) [1] 朱家宏. 階層式深度神經網路及其應用, pages 18-20, 2023. [2] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016. [3] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1-9), 2015. [4] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv 2017, arXiv: 1704.04861.2017. [5] Dmitry Retinskiy. (2020). Multi-Label Image Classification with PyTorch. Retrieved from https://learnopencv.com/multi-label-image-classification-with-pytorch(Oct.13,2023) [6] CNN for deep learning | Convolutional neural networks. Retrieved from https://datapeaker.com/en/big--data/cnn-for-deep-learning-convolutional-neural-networks(Oct.13,2023) [7] LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. [8] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6), 84-90. [9] K. Simonyan and A. Zisserman, Very deep convolutional networks for large-scale image recognition, arXiv preprint arXiv: 1409.1556., 2014. [10] Yan, Z., Zhang, H., Piramuthu, R., Jagadeesh, V., DeCoste, D., Di, W., & Yu, Y.. HD-CNN: hierarchical deep convolutional neural networks for large scale visual recognition. In Proceedings of the IEEE international conference on computer vision (pp. 2740-2748). 2015. [11] Xinqi Zhu, Michael Bain. B-CNN: Branch Convolutional Neural Network for Hierarchical Classification. arXiv 2017, arXiv:1709.09890.2017. [12] Salma Taoufiq , Balázs Nagy , Csaba Benedek.HierarchyNet: Hierarchical CNN-Based Urban Building Classification. Remote Sensing,Volume 12 ,Issue 22 .2020. [13] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Computer Science Department, University of Toronto, Tech. Rep, 2009. [14] D. Arthur and S. Vassilvitskii. K-means++: The advantages of careful seeding. In SODA, pages 1027–1035, 2007. zh_TW
