Research On Methods Of Subspace Feature Extraction Based On Manifold Learning In Face Recognition

Feature extraction is one of the most elementary problems in the area of pattern recognition. It is a critical factor to extract effectively discriminant features for face recognition. As face space is always believed to be a low-dimensional submanifold embedded in high-dimensional ambient space, many methods of subspace feature extraction adopt manifold learning to explore intrinsic structure of face patterns. This dissertation mainly deals with methods of subspace feature extraction based on manifold learning, and the main work and contributions are presented in the following aspects:1. Linear local tangent space alignment does not make use of class information of face samples and the extracted features are redundant, moreover, it can not maintain the measure of high-dimensional data space. Discriminant linear local tangent space alignment using ridge regression is proposed, which takes advantage of class information and spectral regression theory. On this basis, orthogonal discriminant linear local tangent space alignment is proposed with introduction of orthogonalization, which reduces algorithm complexity and maintains the measure.2. When the number of facial feature dimensions is much larger than the number of face samples, the subspace learned from the common linear subspace is not smooth and sparse solutions can not be obtained. To address this situation, a algorithm named sparsely and smoothly marginal fisher analysis is proposed, which adopots spatially smoothly and sparsely regularized techniques and chooses the vertices of regular simplex as the mapped targets, then uses elastic net to construct mapped relationship, at last the smallest feature subset which reflects the best information can be obtained.3. Locality sensitive discriminant analysis can only handle vector-based data, which can not preserve spatial information of image pixels and easily leads to singular problem. Tensor local discriminant projection is proposed to resolve the case, and the actual computation of the algorithm is reduced to a generalized eigenvalue problem to gain two transformation matrices of tensor subspace. The proposed algorithm eliminates the relevance of rows and columns from horizontal and vertical directions, compresses feature dimensions and preserves the integrity of spatial information of the images.4. When facial features are highly nonlinear distributed, linear subspace methods are difficult to extract effectively facial features. Combing the idea of kernel mapping and neighborhood preserving maximal margin analysis, kernel neighborhood preserving maximal margin analysis is proposed, then neighborhood preserving maximal margin analysis using kernel ridge regression is proposed which overcomes the difficulty of extracting nonlinear features and preserves information of geometrical and discriminant structure of face manifold.In the framework of graph embedding, it shows that the proposed algorithms have correspondingly extensional type of graph embedding and applicability.