眼瞼下垂為眼部醫學常見之疾病，其成因為提眼瞼肌功能退化或苗勒氏肌失調所致，造成上眼瞼下垂進而遮蔽視野，因此導致視線範圍受限等問題，除此之外，於心理層面也有所影響，主因為外表改變且不美觀之緣故，甚至造成社交恐懼、焦慮，進而影響生活品質，導致心理問題。因此若能盡早手術並改善便能增進生活品質。
然而評估眼瞼下垂時需要量測多個參數及多種流程。傳統上使用量尺測量，該方法存在主觀性之誤差，且測量過程繁瑣，不僅耗費人力與時間成本，其可靠度更存在一定程度缺陷。為解決上述問題，近年已有相關研究使用深度學習之自動化檢測手段降低人工測量造成之誤差。但該研究忽略病人眼瞼功能存在缺陷且無法自主控制眼瞼活動之病況，導致無法實際應用於診療當中。因此，本研究旨在改善過去已知缺陷，並加以調整系統以提高準確性；不僅能精準檢測眼瞼下垂問題外，同時提供動態眼球追蹤功能。此外，本研究提出更靈活之系統架構，使其可應用於無拘束環境中，不僅便於診間醫師使用，同時提供民眾於日常生活進行檢測。
過去研究利用深度學習訓練語義分割模型以用於檢測眼瞼下垂嚴重性並推導多個參數，因此本研究為取得眼睛輪廓以便進行推導，選擇保留虹鞏膜語義分割模型。但有別於過去，本研究捨棄使用固定機台，旨在落實無拘束環境亦能檢測之目的。透過動態影片評估並實現一更精準且更靈活之檢測系統。本研究將使用視線估計模型－L2CS Net推算眼球移動極值、雙眼同步率及眼球移動速度等。並設計多種演算法校正及調整影像以推算更為精準之評估結果並轉換其公制單位。
最後，本研究對系統進行精準度測試，統計各參數間誤差範圍，旨在驗證本系統於評估眼瞼下垂時之準確性及穩定性，結果表明在MRD1、MRD2、PFH、PFL、LF及Blepharoptosis Severity等六個參數間最小誤差為0至0.03毫米。綜上所述，本研究不僅改良過去系統已知缺陷，並有望落實於診療及日常檢驗。

Blepharoptosis, a common condition in ophthalmology, is caused by the degeneration of the levator palpebrae muscle or dysfunction of the Müller muscle, leading to the drooping of the upper eyelid and obstructing the field of vision. This not only causes visual obstruction and reduces the range of sight but also affects individuals psychologically due to changes in appearance, leading to social anxiety and ultimately impacting quality of life, resulting in psychological issues. Therefore, early surgical intervention and improvement can enhance quality of life.
However, assessing the severity of blepharoptosis requires multiple procedures and tests. Traditional measurements using rulers suffer from subjectivity and are cumbersome, consuming manpower and time, with reliability being somewhat compromised. To solve the above problems, recent research has explored the use of deep learning for automated detection to reduce errors associated with manual measurements. However, these studies overlook cases where patients have defective eyelid function and cannot control eyelid movement autonomously, rendering them impractical for clinical application. Therefore, this study aims to address previously identified shortcomings by enhancing the system to improve accuracy. The improved system not only enables precise detection of ptosis but also incorporates dynamic eye-tracking functionality. Furthermore, this study proposes a more flexible system architecture that can be applied in non-restrictive environments, facilitating use by clinicians in medical settings and allowing the public to conduct self-assessments in daily life.
Previous studies employed deep learning to train semantic segmentation model for detecting the severity of blepharoptosis and deriving multiple parameters. Therefore, in this study, derivation is performed after obtaining the eye contour, hence opting to retain the iris segmentation model. Unlike previous approaches, this study abandons the use of fixed equipment, aiming to achieve unconstrained testing. Through dynamic video assessment, a more precise and flexible detection system is realized. This study will utilize the gaze estimation model "L2CS Net" to estimate eyes movement maximum value, eyes synchronization and eyes motion speed, and design various algorithms to derive more accurate evaluation results and convert them into metric units.
In conclusion, this study conducted an accuracy test on the system and analyzed the error range among various parameters. The aim was to verify the accuracy and stability of the system in evaluating ptosis. The results showed that the minimum error among the six parameters—MRD1, MRD2, PFH, PFL, LF, and Blepharoptosis Severity—ranged from 0 to 0.03 millimeters. In conclusion, this study not only improves upon the known deficiencies of previous systems but also holds promise for implementation in clinical settings and routine examinations.

論文審定書 i
中文摘要 ii
Abstract iii
目錄 v
圖次 vii
表次 ix
第一章 緒論 1
1.1研究背景 1
1.2研究動機 3
1.3研究目的 5
第二章 相關研究 6
2.1機器學習 6
2.1.1監督式學習 6
2.1.2非監督式學習 8
2.1.3半監督式學習 8
2.1.4強化學習 9
2.2深度學習 9
2.3卷積神經網路 10
2.4語義分割 11
2.5 基於語義分割之虹膜鞏膜分割模型 12
2.5.1 模型訓練 12
2.5.2 訓練結果 13
2.6 L2CS Net 14
2.6.1 模型架構 14
2.6.2 訓練結果 15
第三章 研究方法 17
3.1 研究流程與系統架構 17
3.1.1 研究流程 17
3.1.2 系統架構 19
3.2 人臉對齊演算法 20
3.3 視線估計及挑選演算法 22
3.3.1視線估計(Gaze Estimation) 22
3.3.2視線挑選演算法 23
3.4 影像校正 25
3.5眼瞼下垂參數推導方法 27
3.5.1參數推導 28
3.5.2 公制單位換算方法 29
3.6動態眼球評估方法 30
3.7精準度驗證方法 32
3.7.1 FFHQ-UV 32
3.7.2 動態調整眼瞼座標 33
第四章 研究結果 35
4.1精準度驗證結果 35
4.2系統功能 44
4.3 系統結果 46
4.3.1 眼瞼下垂評估結果 47
4.3.2 動態追蹤評估結果 49
第五章 結論 50
參考文獻 51

[1] C.-S. Lai, and Y.-W. Lai, Correction of Blepharoptosis, Aesthetic Plastic Surgery in Asians: Principles and Techniques, Vol. 2, ISBN: 978-1482240870, New York: CRC Press, 2015.

[2] J. Bacharach, W. W. Lee, A. R. Harrison, and T. F. Freddo, "A Review of Acquired Blepharoptosis: Prevalence, Diagnosis, and Current Treatment Options," Eye (London, England), Vol. 35, pp. 2468–2481, 2021.

[3] H. Richards, E. Jenkinson, N. Rumsey, P. White, H. Garrott, H. Herbert, F. Kalapesi, and R. A. Harrad, "The Psychological Well-Being and Appearance Concerns of Patients Presenting with Ptosis," Eye (London, England), Vol. 28, pp. 296–302, 2014.

[4] H.-T. Lai, S.-F. Weng, C.-H. Chang, S.-H. Huang, S.-S. Lee, K.-P. Chang, and C.-S. Lai, "Analysis of Levator Function and Ptosis Severity in Involutional Blepharoptosis," Annals of Plastic Surgery, Vol. 78, pp. S58–S60, 2017.

[5] T. R. N, and R. Gupta, "A Survey on Machine Learning Approaches and Its Techniques," IEEE International Students' Conference on Electrical, Electronics and Computer Science (SCEECS), Bhopal, India, pp. 1–6, 2020.

[6] W. McCulloch, and W. Pitts, "A Logical Calculus of Ideas Immanent in Nervous Activity," Bulletin of Mathematical Biophysics, Vol. 5, Issue 4, pp. 115–133, 1943.

[7] Y.-N. Dong, and G.-S. Liang, "Research and Discussion on Image Recognition and Classification Algorithm Based on Deep Learning," International Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI), Taiyuan, China, pp. 274–278, 2019.

[8] V. Badrinarayanan, A. Kendall, and R. Cipolla, "SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation," IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 39, pp. 2481–2495, 2015.

[9] 黃翰俞, "基於語義分割之自動眼瞼下垂檢測系統," 國立中山大學碩士論文, 二〇二三。

[10] C. Lugaresi, J. Tang, H. Nash, C. McClanahan, E. Uboweja, M. Hays, F. Zhang, C.-L. Chang, M. G. Yong, J. Lee, W.-T. Chang, W. Hua, M. Georg, and M. Grundmann, "Mediapipe: A Framework for Building Perception Pipelines," arXiv:1906.08172, 2019.

[11] A. A. Abdelrahman, T. Hempel, A. Khalifa, A. Al-Hamadi, and L. Dinges, "L2CS-Net: Fine-Grained Gaze Estimation in Unconstrained Environments," 2023 8th International Conference on Frontiers of Signal Processing (ICFSP), Corfu, Greece, pp. 98–102, 2023.

[12] P. Kellnhofer, A. Recasens, S. Stent, W. Matusik, and A. Torralba, "Gaze360: Physically Unconstrained Gaze Estimation in the Wild," IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South), pp. 6911–6920, 2019.

[13] X. Zhang, Y. Sugano, M. Fritz, and A. Bulling, "It’s Written All Over Your Face: Full-Face Appearance-Based Gaze Estimation," Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 2299–2308, 2017.

[14] M. Grinberg, Flask Web Development, 2nd Edition, O'Reilly Media, Inc., 2018.

[15] A. Lambora, K. Gupta, and K. Chopra, "Genetic Algorithm A Literature Review," International Conference on Machine Learning, Big Data, Cloud and Parallel Computing (COMITCon), Faridabad, India, pp. 380–384, 2019.

[16] H. Bai, D. Kang, H. Zhang, J. Pan, and L. Bao, "FFHQ-UV: Normalized Facial UV-Texture Dataset for 3D Face Reconstruction," Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 362–371, 2023.