臺灣博碩士論文加值系統

臉部超解析法被廣泛地應用於重建毀損或失真影像，一般而言，影像的失真是從高解析影像的模糊或像素化，該合成方法可用於重建毀損影像，透過相似標準來近似高解析影像，在本文中，我們將提出一種應用於臉部影像的超解析方法，所提出的方法考慮了會使影像有不同光影差異的各種光源條件。
在提出的方法中，以新的局部特徵強化合成來達到更好的辨識度以及重建效果，同時也使用亮度正規化與單邊直方圖等化這類的前/後處理以提高合成品質，在未來，新的方法可以更進一步應用於極差的臉部影像重建並且改善臉部辨識，為了檢驗新方法的有效性，以亞洲人與西方人臉部影像兩組資料庫個別進行實驗，我們可以很容易地發現模擬結果能達到讓人滿意的效果。

Face hallucination using the technique of super-resolution has been widely adopted to reconstructthe damaged or distorted images. In general, the image distortion is usually blur or pixelized from the original high-resolution.The synthetic method can be used to reconstruct the damaged image to approximate the high-resolution image by specific similar criteria. In this thesis, we will propose a face super-resolution method for face image hallucination. The proposed method considered various conditions of light source that makes images with different lighting appearance.
In proposed method, the new local features is developed to enhance the synthesis and then achieves better identification and reconstruction results. At the same time, the pre/post processing such as light normalization and single side histogram equalization are also developed to improve the synthesis quality. In the future, new method can be further applied to extremely poor face image reconstruction and improve the face recognition. In order to examine the effectiveness of new method, the experience is conducted by two datasets, Asians and Westerners face database, we can easily find that the simulation results achieve satisfy performance.

摘要III
誌謝V
目錄VI
圖目錄IX
一、研究動機與背景1
二、相關性研究回顧3
2.1 全域為基礎的方法(global based methods)6
2.2 區塊為基礎的方法(patch based methods)6
2.3 低解析樣本的生成運算方法9
2.4 以區塊為基礎進行11
2.5 預先重新定義(Redefined Nonlocal Prior)12
2.6 區塊重疊(Overlapped)12
三、區塊鄰近嵌入法(Neighbor Embedding,NE)14
3.1.1 區塊分割16
3.1.2 位置區塊的差異值17
3.1.3 K-NN演算法(K-Nearest Neighbors algorithm)18
3.1.4 權重值計算19
3.1.5 合成20
3.1.6 非局部預先重新定義RNP(Redefined Nonlocal Prior)20
3.1.7 實驗結果23
3.2 全域及局部特徵影像模擬合成26
3.2.1 區塊分割29
3.2.1.1 並排的區塊分割29
3.2.1.2 局部特徵區塊分割30
3.2.2 亮度調整32
3.2.2.1 亮度正規化32
3.2.2.2 區塊亮度校正樣本36
3.2.2.3 失真影像的影響38
3.2.2.4 特徵區塊正規化進行NE法40
3.2.3 全域與區塊亮度校正NE法結果比較42
3.2.4 非特徵域與臉部特徵細節43
3.2.4.1 非特徵43
3.2.4.2 單邊直方圖等化44
3.2.5 模擬合成48
四、實驗及討論53
4.1 人臉資料庫54
4.2 實驗測試結論68
五、結論與未來研究方向69
六、參考文獻71

[1]Baker S, Kanade T. Limits on super-resolution and how to break them[J]. Pattern Anal-ysis and Machine Intelligence, IEEE Transactions on, 2002, 24(9): 1167-1183.
[2]Baker, S. and Kanade, T. 2000a. Hallucinating faces. In IEEE International Conference on Automatic Face and Gesture Recognition.
[3] Wang X, Tang X. Hallucinating face by eigentransformation[J]. Systems, Man, and Cy-bernetics, Part C: Applications and Reviews, IEEE Transactions on, 2005, 35(3): 425-434.
[4]Chang H, Yeung D Y, Xiong Y. Super-resolution through neighbor embed-ding[C]//Computer Vision and Pattern Recognition, 2004. CVPR 2004. Proceedings of the 2004 IEEE Computer Society Conference on. IEEE, 2004, 1: I-I.
[5]Ma X, Zhang J, Qi C. Position-based face hallucination method[C]//Multimedia and Expo, 2009. ICME 2009. IEEE International Conference on. IEEE, 2009: 290-293.
[6]Liang Chen Ruimin Hu. Face Super resolution Based on Rede_ned NonlocalPrior for VLQ Scenarios.
[7]Yueting Zhuang, Jian Zhang∗, FeiWu. Hallucinating faces: LPH super-resolution and neighbor reconstruction for residue compensation. 2007
[8]Ce Liu∗, Heung-Yeung Shum, William t. Freeman Face Hallucination: Theory and Practice, 2007
[9]Guan-da Huang. The Study of Effective Texture Synthesis and Image Feature Transfer, 2009
[10]Chuang, Chialung. Piece-Wise Histogram Equalization For Image Enhancement, 2012