臺灣博碩士論文加值系統

在實際應用中，神經網路的泛化能力常常不足。我們觀察到，當測試圖像與訓練所使用的圖像在如角度、顏色、距離等場域上的變化時，訓練出的網路往往無法在測試圖像中取得優異的分類結果。此外，對每一張訓練圖像進行分類標註往往曠日費時，尤其當需要標註的圖像涉及高專業技術時，這將為整個任務帶來大量的金錢和時間成本。為了解決這類問題，本文提出了一種新深度學習模型並稱為條件共用特徵擷取網路(CSFEN)，旨在有效利用現有的跨領域圖像資料進行半監督式學習。此網路嘗試找出含標籤圖像資料與不同領域未含標籤圖像資料的共用特徵，並結合領域適應方法進行分類。為了驗證此網路的有效性，本論文從多個方面進行分析，例如當未標註圖像數量大於含標籤圖像數量時，以及當未標註圖像與含標籤圖像在顏色、風格、角度等方面存在差異時。此網路在多個不同類型的資料集中已取得了優異的分類效果。相比於過去的方法，所提的網路展現了相對學習穩定性、解釋性以及具競爭性的分類表現。

In practical applications, the generalization ability of neural networks is frequently inadequate. It has been observed that when there are domain changes such as angles, colors, and distances between test images and the images used for training, a trained network often fails to achieve excellent classification results in the test images. Furthermore, the process of classifying and labelling each training image is often time-consuming, particularly when the images in question require a high level of expertise, which can result in significant costs in terms of both time and financial resources. To address these issues, this thesis presents a Conditional Shared Feature Extraction Network (CSFEN) that aims to optimize the utilization of existing data in different domains to perform semi-supervised learning. The objective is to identify the shared features of labelled and unlabeled image data from different domains and combine them with a domain adaptation method for classification. In order to verify the effectiveness of the proposed network, this thesis analyzes the data from different perspectives. These include situations where the number of unlabeled images exceeds that of the labelled images, as well as instances where there are discrepancies between the unlabeled and labelled images with regard to factors such as color, style, and angle. The proposed network shows excellent classification results across a diverse range of image datasets. In comparison to previous methods, the proposed network exhibits superior learning stability, interpretability, and competitive classification results.

摘要 i
Abstract ii
Contents iii
List of Figures v
List of Tables viii
Chapter 1 Introduction 1
1.1 Convolution Neural Network based Classifier Survey 1
1.2 Unsupervised Domain Adaptation Method Survey 2
1.3 Contributions of this Thesis 4
1.4 Organization of this Thesis 5
Chapter 2 Materials 7
2.1 Office31 7
2.2 Office-Home 10
2.3 Glomeruli 11
2.4 Human Posture 15
Chapter 3 Domain Adaptation for Classification 19
3.1 Network Structure of SFEN 19
3.2 Network Structure of Xception 20
3.3 Separator Network 21
3.4 Domain Adversarial Network 23
Chapter 4 Learning Algorithm 27
4.1 Forward path of CSFEN 27
4.2 Mixup strategy 27
4.3 Loss Function of CSFEN 29
4.4 Training Detail 32
Chapter 5 Experimental Results 35
5.1 Performance Metrics 35
5.2 Office31 Dataset 36
5.3 Office-Home Dataset 40
5.4 Glomerulus Dataset 42
5.5 Human Posture Dataset 46
5.6 Discussions 48
5.6.1 Ablation Study 48
5.6.2 Stable Training Process 50
Chapter 6 Conclusion 59
References 60

[1]Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, Jan. 1998, doi: 10.1109/5.726791.
[2]S. Liu and W. Deng, “Very deep convolutional neural network based image classification using small training sample size,” in 2015 3rd IAPR Asian Conference on Pattern Recognition (ACPR), Jan. 2015, pp. 730–734. doi: 10.1109/ACPR.2015.7486599.
[3]K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA: IEEE, Jun. 2016, pp. 770–778. doi: 10.1109/CVPR.2016.90.
[4]F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jul. 2017, pp. 1800–1807. doi: 10.1109/CVPR.2017.195.
[5]S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan, “A theory of learning from different domains,” Mach Learn, vol. 79, no. 1, pp. 151–175, May 2010, doi: 10.1007/s10994-009-5152-4.
[6]Y. Ganin et al., “Domain-adversarial training of neural networks,” in Domain Adaptation in Computer Vision Applications, G. Csurka, Ed., in Advances in Computer Vision and Pattern Recognition. , Cham: Springer International Publishing, 2017, pp. 189–209. doi: 10.1007/978-3-319-58347-1_10.
[7]E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell, “Adversarial discriminative domain adaptation,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI: IEEE, Jul. 2017, pp. 2962–2971. doi: 10.1109/CVPR.2017.316.
[8]M. Long, Z. CAO, J. Wang, and M. I. Jordan, “Conditional adversarial domain adaptation,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2018. Accessed: Aug. 06, 2024. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2018/hash/ab88b15733f543179858600245108dd8-Abstract.html
[9]K. Bousmalis, G. Trigeorgis, N. Silberman, D. Krishnan, and D. Erhan, “Domain separation networks,” Aug. 21, 2016, arXiv: arXiv:1608.06019. doi: 10.48550/arXiv.1608.06019.
[10]A. Sharma, T. Kalluri, and M. Chandraker, “Instance level affinity-based transfer for unsupervised domain adaptation,” Apr. 02, 2021, arXiv: arXiv:2104.01286. doi: 10.48550/arXiv.2104.01286.
[11]N. Xiao and L. Zhang, “Dynamic weighted learning for unsupervised domain adaptation,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA: IEEE, Jun. 2021, pp. 15237–15246. doi: 10.1109/CVPR46437.2021.01499.
[12]J. Na, H. Jung, H. J. Chang, and W. Hwang, “FixBi: Bridging domain spaces for unsupervised domain adaptation,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2021, pp. 1094–1103. doi: 10.1109/CVPR46437.2021.00115.
[13]T. Westfechtel, H.-W. Yeh, D. Zhang, and T. Harada, “Gradual source domain expansion for unsupervised domain adaptation,” in 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Waikoloa, HI, USA: IEEE, Jan. 2024, pp. 1935–1944. doi: 10.1109/WACV57701.2024.00195.
[14]S. Lee, S. Cho, and S. Im, “DRANet: Disentangling representation and adaptation networks for unsupervised cross-domain adaptation,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA: IEEE, Jun. 2021, pp. 15247–15256. doi: 10.1109/CVPR46437.2021.01500.
[15]H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk minimization,” Apr. 27, 2018, arXiv: arXiv:1710.09412. doi: 10.48550/arXiv.1710.09412.
[16]M. Xu et al., “Adversarial domain adaptation with domain mixup,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 04, Art. no. 04, Apr. 2020, doi: 10.1609/aaai.v34i04.6123.
[17]Y. Wu, D. Inkpen, and A. El-Roby, “Dual mixup regularized learning for adversarial domain adaptation,” in Computer Vision – ECCV 2020, vol. 12374, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds., in Lecture Notes in Computer Science, vol. 12374. , Cham: Springer International Publishing, 2020, pp. 540–555. doi: 10.1007/978-3-030-58526-6_32.
[18]S. Yan, H. Song, N. Li, L. Zou, and L. Ren, “Improve unsupervised domain adaptation with mixup training,” Jan. 02, 2020, arXiv: arXiv:2001.00677. doi: 10.48550/arXiv.2001.00677.
[19]J. Na, D. Han, H. J. Chang, and W. Hwang, “Contrastive vicinal space for unsupervised domain adaptation,” in Computer Vision – ECCV 2022, vol. 13694, S. Avidan, G. Brostow, M. Cissé, G. M. Farinella, and T. Hassner, Eds., in Lecture Notes in Computer Science, vol. 13694. , Cham: Springer Nature Switzerland, 2022, pp. 92–110. doi: 10.1007/978-3-031-19830-4_6.
[20]V. Gliner, V. Makarov, A. I. Avetisyan, A. Schuster, and Y. Yaniv, “Using domain adaptation for classification of healthy and disease conditions from mobile-captured images of standard 12-lead electrocardiograms,” Sci Rep, vol. 13, no. 1, p. 14023, Aug. 2023, doi: 10.1038/s41598-023-40693-6.
[21]Y.-C. Lo, I.-F. Chung, S.-N. Guo, M.-C. Wen, and C.-F. Juang, “Cycle-consistent GAN-based stain translation of renal pathology images with glomerulus detection application,” Applied Soft Computing, vol. 98, p. 106822, Jan. 2021, doi: 10.1016/j.asoc.2020.106822.
[22]J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using Cycle-Consistent adversarial networks,” in 2017 IEEE International Conference on Computer Vision (ICCV), Venice: IEEE, Oct. 2017, pp. 2242–2251. doi: 10.1109/ICCV.2017.244.
[23]C.-F. Juang, Y.-W. Chuang, G.-W. Lin, I.-F. Chung, and Y.-C. Lo, “Deep learning-based glomerulus detection and classification with generative morphology augmentation in renal pathology images,” Computerized Medical Imaging and Graphics, vol. 115, p. 102375, Jul. 2024, doi: 10.1016/j.compmedimag.2024.102375.
[24]X. Long, J. Liu, and X. Hou, “Domain adaptation of digital pathology images using joint stain color and image quality constraints,” in 2023 IEEE International Conference on Image Processing (ICIP), Kuala Lumpur, Malaysia: IEEE, Oct. 2023, pp. 1805–1809. doi: 10.1109/ICIP49359.2023.10222270.
[25]E. Othman, Y. Bazi, F. Melgani, H. Alhichri, N. Alajlan, and M. Zuair, “Domain adaptation network for cross-scene classification,” IEEE Trans. Geosci. Remote Sensing, vol. 55, no. 8, pp. 4441–4456, Aug. 2017, doi: 10.1109/TGRS.2017.2692281.
[26]Z. Zheng, Y. Zhong, Y. Su, and A. Ma, “Domain adaptation via a task-specific classifier framework for remote sensing cross-scene classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–13, 2022, doi: 10.1109/TGRS.2022.3151689.
[27]R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” in 2017 IEEE International Conference on Computer Vision (ICCV), Oct. 2017, pp. 618–626. doi: 10.1109/ICCV.2017.74.
[28]D. Hutchison et al., “Adapting visual category models to new domains,” in Computer Vision – ECCV 2010, vol. 6314, K. Daniilidis, P. Maragos, and N. Paragios, Eds., in Lecture Notes in Computer Science, vol. 6314. , Berlin, Heidelberg: Springer Berlin Heidelberg, 2010, pp. 213–226. doi: 10.1007/978-3-642-15561-1_16.
[29]H. Venkateswara, J. Eusebio, S. Chakraborty, and S. Panchanathan, “Deep hashing network for unsupervised domain adaptation”.
[30]M. Mirza and S. Osindero, “Conditional generative adversarial nets,” Nov. 06, 2014, arXiv: arXiv:1411.1784. doi: 10.48550/arXiv.1411.1784.
[31]Y. Grandvalet and Y. Bengio, “Semi-supervised learning by entropy minimization,” in Advances in Neural Information Processing Systems, MIT Press, 2004. Accessed: Aug. 07, 2024. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2004/hash/96f2b50b5d3613adf9c27049b2a888c7-Abstract.html
[32]S. Li et al., “Learning to reconstruct crack profiles for eddy current nondestructive testing,” Oct. 28, 2019, arXiv: arXiv:1910.08721. doi: 10.48550/arXiv.1910.08721.
[33]L. Liu et al., “On the variance of the adaptive learning rate and beyond,” Oct. 25, 2021, arXiv: arXiv:1908.03265. doi: 10.48550/arXiv.1908.03265.
[34]M. R. Zhang, J. Lucas, G. Hinton, and J. Ba, “Lookahead optimizer: k steps forward, 1 step back,” Dec. 03, 2019, arXiv: arXiv:1907.08610. doi: 10.48550/arXiv.1907.08610.
[35]I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” May 03, 2017, arXiv: arXiv:1608.03983. Accessed: Aug. 08, 2024. [Online]. Available: http://arxiv.org/abs/1608.03983
[36]J. Zhu, H. Bai, and L. Wang, “Patch-Mix transformer for unsupervised domain adaptation: A game perspective,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada: IEEE, Jun. 2023, pp. 3561–3571. doi: 10.1109/CVPR52729.2023.00347.