一般來說，人臉影像光影變化會嚴重地影響到人臉辨識系統的辨識率。為了減少光照條件的影響，我們提出以樣本學習的方式為基礎的光線轉換技術。系統的目的是將測試影像透過CVAE模型轉換輸入影像的光照條件，目的是使光線條件與資料庫比對樣本影像的光照條件一致，並藉此提高辨識率。實驗結果顯示，我們的方法不需要手動標記人臉影像的陰影區域，給定一張測試人臉影像與一群人臉訓練資料集，系統會自動改變測試影像的光照條件。

In general, face recognition systems are not robust against illumination variation; in this study, we propose a learning-based framework to reconstruct a face in different lighting conditions in order to improve the performance of downstream applications such as automatic face recognition. The proposed framework is based on conditional variational autoencoder (CVAE) network to disentangle the identity and the shadow effect. Our proposed framework doesn’t need to label the shadow area of faces in any way. It is trained with a loss function to better preserve the identity and reconstruction result. YaleB face database with illumination variation are used for training and evaluation purpose. The experimental results show that the proposed framework can automatically remove shadows from a single image.

目錄
第一章 緒論	1
1.1 研究動機與目的	1
第二章 相關研究	4
第三章 系統架構	6
3.1 資料章節	6
3.2 條件式變分自動編碼機 CVAE	7
3.3人臉鑑別器	9
3.4 W生成器模型	9
第四章 實驗結果與分析	13
4.1 實驗設定	13
4.2 CVAE 光線轉換	13
4.3 辨識率	14
第五章 結論	16
參考文獻	17
附錄: 英文論文	19

圖目錄
圖 1 CVAE 加條件方式示意圖	5
圖 2 系統架構圖	6
圖 3(a) Extended YaleB中挑選一人顯示資料庫中的不同角度光源的人臉影像;(b) 正面光源(光源角度編號+000e+00)人臉影像範例;(c) 左下光源(光源角度編號+035e-20)人臉影像範例;(d) 右側光源(光源角度編號-050e+00)人臉影像範例	7
圖 4 側光轉正光CVAE結果	9
圖 5 只使用Lcontent的結果	11
圖 6 Lcontent+LID 結果	12
圖 7 CVAE 5種側光轉正光結果	14
 
表目錄
表格 1 人臉驗證器在11種光線下與正面光影像比對的辨識率	15
表格 2 本篇方法重建影像的辨識率	15

[1] F. Zhao, J. Feng, J. Zhao, W. Yang, and S. Yan. Robust lstm-autoencoders for face de-occlusion. In the wild.arXiv:1612.08534, 2016.
[2] S. Chopra, R. Hadsell, and Y. LeCun. Learning a similarity metric discriminatively, with application to face verification. In CVPR, 2005
[3] D. P. Kingma and M. Welling. Auto-encoding variational Bayes. In ICLR, 2013.
[4] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, 2008.
[5] A. Georghiades, P. Belhumeur, and D. Kriegman, “From few to many: Illumination cone models for face recognition under variable lighting and pose,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 23, no. 6, pp. 643–660, Jun. 2001.
[6] R. A. Yeh, C. Chen, T. Y. Lim, A. G. Schwing, M. HasegawaJohnson, and M. N. Do. Semantic image inpainting with deep generative models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 5485–5493, 2017.
[7] G. Liu, F. A. Reda, K. J. Shih, T.-C. Wang, A. Tao, and B. Catanzaro.
20
Image inpainting for irregular holes using partial convolutions. arXiv preprint arXiv:1804.07723, 2018.
[8] D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros. Context encoders: Feature learning by inpainting. CVPR, 2016.
[9] K. Sohn, H. Lee, and X. Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems, pages 3483–3491, 2015.