Face Recognition Based On Manifold Learning

As the development of science and technology, information usually comes with high dimensional data. However, people's perceptions are only limited in three dimensionalities and dimensional reduction becomes an indispensable and an effective analysis method to deal with high dimensional data. Compared with old linear dimensional reduction algorithm, Manifold learning is a nonlinear dimensional reduction method that can perverse some character from high dimensionality to low dimensionality. Face recognition is still a hot research problem and people try to set a model to approach the real face images space. Manifold learning learns the rule of the real face images space, which can present the geometric character better than European space. Manifold learning learns global or local character which still exist some problems, such as parameters determination, computational complexity. This paper focused on the unsupervised limitation. The main contributions are as follows:1. ISOMAP, as an unsupervised manifold algorithm, does not use the class information of high dimensional data and the Extended ISOMAP combines Linear Discriminant Analysis(LDA, for short), a linear classification, which could not performance well in present the nonlinear of data inside structure, with traditional ISOMAP. This paper introduces an adaptive distance factor, and makes full use of the given class information, then achieved nonlinear classification and experiments identify the effective of the Adaptive Distance ISOMAP (ADI, for short) algorithm. Then, considering the computational complexity of ADI, we combined ADI with Generalization of ISOMAP(GI, for short) and proposed ADGI which can be used in real time.2. Traditional Locally Linear Embedding(LLE, for short) algorithm chooses sample's neighborhood simply, so it may not work well when high dimension data already has class information. Considering the location relationship between sample point and center point of every class, we propose a hierarchical method to choose the neighborhood, which can exactly get the neighborhood of sample point.