Research On Algorithms Of Face Recognition Based On Semi-supervised Manifold Learning Theory

Human face recognition has become a research focus in pattern recognition, because it is a most development potential biometric technology for the characteristics of non-contact, safe and convenient. However, because of the complex facial structure, the diverse facial expression and the changing light intensity, face recognition is still being recognized as a challenging research. In recent years, more and more research shows that human face is a embedded manifold structure. High dimensional face data is a low-dimensional non-linear manifold formed by some internal controlled variables, such as facial expression, illumination condition and face pose. If some controlled variables can be found, observation space dimensionality could be compressed significantly. Not only avoid the so-called "dimension disaster", but also improve the recognition accuracy.In this dissertation, based on the theory of manifold learning, we has researched on two face recognition algorithms of extracting features and their effectiveness has been tested by experimental results. Its main research contents and the achievements are as follows:1. A exhaustive analysis is made on four manifold learning representative algorithm including Isometric Feature Mapping, locally linear embedding, Laplacian eignmap, and local tangent space alignment algorithm, then analysed the aims, principles and process of solving, specifically estimated the time complexity of the algorithms. Finally, we summarized the framework of manifold learning algorithms.2. A analysis is made on some key issues such as dimensionality reduction, linear manifold learning, supervision and semi-supervised manifold learning when manifold learning is applied to face recognition. Based on this analysis, this dissertation reveal the the differences and relations of this some key issues.3. A semi-supervised manifold learning algorithm of face recognition is proposed, which is based on the Unsupervised Discriminant Projection. On the basis of the character of the face image, this method gets K-nearest neighbors of each sample by calculating the image euclidean distance, and the adjacency matrix of unsupervised discriminant projection was modified accordingly. Finally, the proposed method that combined labeled samples with modified unsupervised discriminant projection is presented to achieve optimal geometric projection. Extensive experimental results on several public face database validate the correctness and effectiveness of the proposed approach. Meanwhile, we get the distance between every two face images by the image Euclidean distance, and by mathematical analysis it is proved that the image Euclidean distance is able to reflect the different structural information between face images well.4. A kind of semi-supervised manifold learning algorithm of image recognition is presented, which is based on the Neighborhood Preserving Projections. Aiming at the character of the face image, this method gets k-nearest neighbors of each sample by calculating the image euclidean distance, and the reconstruction weight matrix of neighborhood preserving projections was modified accordingly. Finally, the proposed method achieves optimal geometric projection while combining both local structure information and labeled samples by get the identification information from Nonparametric Discriminant Analysis. In the experiments of three pose-varied face database, comparing with the other methods, it is proved that this algorithm is able to enhance the accuracy and stability of face image classification further.