李秀惠Lee, Hsiu-Hui臺灣大學:資訊工程學研究所邱建霖Chiou, Chien-LinChien-LinChiou2010-06-022018-07-052010-06-022018-07-052008U0001-2801200814350200http://ntur.lib.ntu.edu.tw//handle/246246/184917人臉辨識為生物認證中重要的一環。 人臉辨識研究雖然已有時日，但是仍然有許多影響辨識率的因素，如照度(光源不均)、角度、大小、正面角度、表情、遮蔽物等因素。 為了提高人臉辨識率，這些問題必須去避免或者去解決。 近年來有許多學者提出方法來解決這些因素，但尚未有完美的解決方法。 在本篇論文中，我們提出一個照度補償的機制來解決在不同照度下辨識人臉的問題。 另外我們也利用了偵測眼位來解決大小和正面小角度因素，利用正面人臉偵測器來避免了角度和正面大角度的問題。 我們的人臉辨識系統主要包含了四個階段： （1)人臉截取 （2)人臉正規化 （3)人臉特徵截取（4)人臉分類模型建立/預測。 在人臉截取階段，我們利用了 AdaBoost演算法 以及 皮膚色偵測演算法在影像序列中截取出正面人臉來當作訓練樣本集/測試樣本集。另外在進入特徵截取階段之前對樣本集做了正規化(前置處理)。 包括了利用眼位偵測來做到臉部大小以及正面角度的正規化、利用我們提出的方法來補償照度問題。 至於特徵截取階段和分類模型建立/預測階段，我們利用主成份分析 (PCA)先對人臉樣本做降維，再利用SVM 來訓練出人臉分類模型/預測未知人臉。 我們利用了耶魯大學所創建的公用人臉資料庫(Yale Face Database B)來評比我們提出的光照補償方法。 實驗結果顯示，我們的光照補償方法在光照較艱難以及訊練集光源不足的情況下，皆能獲得比其他方法更高的辨識率。Face recognition is one important topic of biometric recognition. Face recognition has been researched for a long time, but there are kinds of factors that reduce the recognition rate such as illumination, pose, scale, orientation, facial expression, occlusion, etc. To improve the recognition rate, we need to avoid or to solve these factors. Though many researchers paid their efforts for solving these factors, there are no perfect solutions. In this thesis, we proposed an illumination compensation mechanism for the face recognition system. Furthermore, we solved scale and slight orientation factors by using eye detectors and avoided pose and heavy orientation factors by using a frontal face detector. There are four phases in our face recognition system: (1) face extraction (2) face normalization (3) face feature extraction (4) face classification model construction/ prediction. In face extraction phase, we extracted face by using an AdaBoost-based face detector and a skin-color detector. As for the normalization phase, we normalized three factors (scale, orientation, and illumination). We first detected the eyes for normalizing scale and orientation. Then we compensated illumination by our illumination compensation method. In feature extraction phase, we reduced the dimension by using PCA. In model construction/prediction phase, we trained the face classification model and predicted unknown face by using SVM. Finally, we evaluated our proposed illumination compensation method by the face databases, Yale Face Database B and Extended Yale Face Database B. And the experiment shows that our method yields higher face recognition rate than other methods under hard illumination conditions even when the lighting condition of training set is deficient.中文摘要 vbstract vihapter 1 Introduction 1.1 Object/Face Recognition System Overview 1.2 Motivation 3.3 Research Goal 6.4 Related Works 6.5 Organization of this Thesis 8hapter 2 Background 9.1 Color Spaces & Transformation 9.1.1 RGB & HSV Color spaces 9.1.2 Conversion between RGB-HSV Color Spaces 11.2 Adaptive Boosting Algorithm (AdaBoost) 13.3 Principal Component Analysis (PCA) 17.3.1 PCA Deviation by Covariance 18.3.2 PCA: Step by Step 19.4 Support Vector Machine (SVM) 21.4.1 SVM Concepts 21.4.2 Non-Linear Classification 25.4.3 SVM for Multi-Class 26.5 Discrete Wavelet Transform (DWT) 27.5.1 The Wavelet Series Expansions 28.5.2 1D Discrete Wavelet Transform 29.5.3 The Fast Wavelet Transform (FWT) 29.5.4 2D Discrete Wavelet Transform 31hapter 3 System Implementation 34.1 System Overview 34.2 Face Extraction Phase 38.2.1 AdaBoost-based Face Detector 39.2.2 Skin Detector under HSV Color Space 42.3 Face Normalization Phase 43.3.1 Scale & Orientation Normalization 43.3.1.1 ROI Setting 44.3.1.2 Eye Detector 44.3.1.3 Rotation and Scaling by Eye Positions 45.3.2 Illumination Normalization 46.3.2.1 Adaptive Brightness & Contrast Adjustment (ABC) 47.3.2.2 2D Discrete Haar Wavelet Transform (DHWT) 49.3.2.3 Log-Mapping and Scaling 49.4 Face Feature Extraction Phase 50.5 Face Model-Construction/Prediction Phase 51hapter 4 Experiment Design & Results 53.1 The Face Database 53.2 Experiment Architecture 56.3 Illumination Compensation Methods 57.4 Experiment Results 64.4.1 Lighting Conditions of Test Set vs. Recognition Rate 65.4.2 The Number of Training Subjects vs. Recognition Rate 69.4.3 Lighting Conditions of Training Set vs. Recognition Rate 71hapter 5 Conclusion and Future Work 73.1 Conclusion 73.2 Future Work 73eferences 75application/pdf4783775 bytesapplication/pdfen-US照度補償人臉辨識人臉偵測皮膚色偵測illumination compensationface recognitionobject detectionskin-color detectionthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/184917/1/ntu-97-R94922099-1.pdf