由於近代對安全控管的需求日益增加,基於生物特徵(Biometrics)進行個人身份認證的技術也越受重視。生物特徵是唯一且專屬於個人的,包含有視網膜、指紋、聲紋和人臉等,以人臉影像做為身份認證的媒介是最可行且方便的方法,會員人臉確認也是研究學者最關切的問題。在會員人臉影像的分類問題上,首要面對的是如何建構龐大且未知的非會員人臉影像資料,或是如何去建構分類器的決策超平面(Decision hyperplane)以將這些未知的非會員人臉影像做正確的分類。本論文提出結合單一類別分類(One-class classification)方法與二元類別分類 (Binary classification)方法所建構而成的分類器。首先,利用主成份分析(Principal Component Analysis, PCA)演算法對人臉影像抽取其最具鑑別力的特徵以縮減資料維度,再使用支持向量資料描述(Support Vector Data Description, SVDD)方法,將會員資料轉換至新的高維特徵空間,於是,一個球狀的決策邊界被學習用來排除大部份的非會員,此外,使用支持向量機器(Support Vector Machines,SVM)建構最佳分離超平面(Optimal Separating Hyperplane, OSH)來區別超球體(Hypersphere)內的會員與非會員。實驗結果顯示,在未經過訓練的非會員人臉增加的情況下,由本論文提出的結合方法所建構的分類器優於SVM、非平衡式支持向量機器(Imbalanced SVM, ISVM)、集成支持向量機器(SVM ensemble)與SVDD分類器,並證明本論文提出的分類器在處理會員人臉確認的問題上具有較佳的穩定性與擴充性。
Owing to the urgent demand of the security system in recent year, the technology related to identity authentication using biometrics has received much attention. Biometrics such as retina, fingerprint, voiceprint and human face of every individual are unique and exclusive. Among these natures, facial feature-based recognition is one of the most straightforward and feasible ways to develop identity authentication systems. For face membership authentication systems, the challenge is how to collect numerous data for the training process, especially the unknown nonmember set. In addition, the construction of the decision hyperplane to accurately classify the unseen nonmember patterns remains difficult. This thesis proposes a new classifier which combines both one-class and binary classification strategies. The principal component analysis (PCA) is utilized to extract the most discriminating features and reduce the dimensionality of input training data first, then employ the support vector data description (SVDD) to map the member data to high dimensional feature space. Then, a spherical decision boundary is learned to exclude most nonmembers. Furthermore, support vector machines (SVM) is used to construct an optimal separating hyperplane (OSH) to distinguish the members and those few nonmembers inside the hypersphere. By using the fusion strategy, the proposed classifier outperforms others, such as SVM, ISVM (Imbalanced SVM), SVM ensemble, SVDD, when faced with nonmembers that are not included in the training. The results also indicate that the proposed classifier is able to gain better stability and generalization performance for face membership authentication.