Manifold-based Feature Extraction And Face Recognition Analysis

Face recognition is an important research field of pattern recognition and has a very wide range application in computer vision, customer identification, multimedia data retrieval and other fields. Feature extraction is the basic problem of face recognition. Studies in recent years have shown that the data of human face is drawn from a non-linear structure distribution. Manifold learning, a kind of new non-linear dimension reduction methods, has attracted a wide range of concerns in recent years, but its application in pattern recognition problems is very difficult. This thesis made an in-depth research on feature extraction, manifold learning and their combination, and developed several effective feature extraction algorithms. The experiment in face image databases proved the effectiveness of improved algorithm. The main works of the thesis include:1. Unsupervised discriminant projection (UDP) was developed for dimensionality reduction of high-dimensional data. It takes account both the local characteristics and nonlocal characteristics. However, the applicability of UDP to high-dimensional image recognition tasks such as face recognition inevitably suffers from the"small sample size"(SSS) problem. It is well-know that maximum scatter difference discriminant analysis essentially avoids the SSS problem using scatter difference discriminant criterion. We propose maximum scatter difference unsupervised discriminant feature extraction that adopts the difference of nonlocal scatter and local scatter. Experiments performed on both ORL and AR face database verify the effectiveness of the proposed method.2. Kernel Locality Preserving Projection (KLPP) is a nonlinear dimension reduction method, which combines the kernel trick with the manifold learning method effectively. It only takes account the local characteristics of samples and neglects the effective information that is important to classification. Although Unsupervised Discriminant Projection (UDP) combines the local characteristics and nonlocal characteristics of samples, it is a linear subspace learning method in nature. It is not able to extract the nonlinear feature of samples. We propose kernel-based unsupervised discriminant projection which takes account both the local characteristics and nonlocal characteristics of samples. It can also describe the nonlinear change of the face effectively. The experiments on Yale face database show that the proposed method outperforms UDP and PCA, and has a good effect in classification.3. Linear discrimination analysis (LDA) is a classical linear dimension reduction method. It is not able to discover the samples'nonlinear structure. However, it is effective on classification through using the type information of samples. Unsupervised discriminant projection (UDP), like most manifold learning algorithms, has not make use of the samples'class information that is useful for classification. We develop a manifold-based supervised feature extraction. It seeks to find a projection that maximizes the nonlocal scatter, while minimizes the local scatter and the within-class scatter. This method can not only find the intrinsic low-dimensional nonlinear data structure, but also is effective on classification. Besides, we combine maximum scatter different criteria with manifold-based supervised feature extraction, which avoids the small sample size problem. The experiments on Yale face image database and ORL face database show the effectiveness of the proposed methods.