Research On Manifold Learning Method Based On Multi-information Fusion

Manifold learning method is a prospective study field, its essence is to find a low-dimensional manifold structure from the high-dimensional samples, and then find the corresponding embedding map to reduce data’s dimension. At present, many manifold learning methods are unsupervised learning methods, these methods can not get satisfactory classification results. Integrated label, global structure and local structure, and other information into the manifold learning methods, can contribute to the improvement of the classification performance.The key to face recognition is to extract the most discriminating features. Face has a low-dimensional manifold structure, so using manifold learning method can obtain better features. Based on manifold learning algorithms, some relative research was made in this paper.The creative work of the thesis includes:1. Fusing the training samples’label information, local structure information, adjacency information, both locality preserving projection algorithm based on repulsion and affinity graphs and graph-based supervised discriminant projection algorithm are proposed.In order to overcome the shortcoming of which Locality Preserving Projections do not use label information for face recognition, this paper proposes a Locality Preserving Projection Algorithm based on Repulsion and Affinity Graphs (LPP-RA) for face recognition. The algorithm constructs repulsion graphs and affinity graphs based on k-NN graphs and combines repulsion graph with affinity graph to extract feature. Repulsion graph reflects the relationship between two near-by samples in different classes and affinity graphs reflect the relationship between two samples which is not near-by, but in the same class. Further more, we define the similarity of samples to remove the effects of noise and eigenvalue variation when extracting the original feature. The experiments on Feret and Yale face image datebase show the effectiveness of the proposed algorithm.Unsupervised Discriminant Projection algorithm is a kind of supervised algorithm, which does not use label information, so it does not use discriminant information of samples. This paper proposes a Graph-based Supervised Discriminant Projection algorithm (GSDP) based on Unsupervised Discriminant Projection algorithm. The algorithm use repulsion graphs and affinity graphs to extract feature. The purpose of using affinity graphs is to make two samples which are in the same class but not nearby attractive and the purpose of constracting repulsion graphs is to repel two samples which are nearby and in different class. The experiments on Feret and Yale face image datebase show the effectiveness of the proposed algorithm.2. Fusing the training samples’local structure information, label information, adjacency information, adaptive neighborhood locality preserving projestion algorithm is proposed.Parameter k affects the performance of the locality preserving projection algorithm, and its difficult is how to choose the ideal parameter k. this paper proposes Adaptive Neighborhood Locality Preserving Projestion (ANLPP) algorithm, whose adaptive neighborhood is not related to the parameter k. When determine the neighbor of the sample, the label information and manifold structure information of samples are used. The constructed weight matrix considers the similality among samples which are from the same class and close to each other, from the same class but not close to each other and from the different class but close to each other. Constructing the objective function by weight matrix, the transformation matrix is obtained. The experiments on Feret and Yale face image datebases show the effectiveness of the proposed algorithm.3. Fusing of the training samples’local structure information, label information, adjacency information, supervised linear local tangent space alignment is proposed.Linear local tangent space alignment algorithm doesn’t use sample’s label information, so this paper proposes linear local tangent space alignment with label information algorithm based on linear local tangent space alignment algorithm (SLLTSA). This algorithm adds label information, and merges together the multiple objective function. It not only keeps original geometrical structure, but also increases discriminant information and also makes the samples from the interclass closer and makes the samples from the intraclass further, so it improves the effect of classification. Compared with some similar algorithms, this algorithm has stronger classification performance.4. Fusing of the training samples’local structure information, interclass and intraclass scatter informations, locality preserving manifold learning algorithm based on Mahalanobis-distance is proposed.Through increasing the Mahalanobis distance of intraclass and reducing the Mahalanobis distance of interclass, the Semisupervised Metric Learning by Maximizing Constraint Margin algorithm can obtain good classification result. However, the algorithm only considers the global structure information of the samples and doesn’t use local structure information of the samples. This paper proposes Locality Preserving Manifold Learning Algorithm based on Mahalanobis-distance (LPMLAM). It not only incrreases the Mahalanobis distance of intraclass and reduces the Mahalanobis distance of interclass, but also keeps samples’local structure information. Yale and Orl face image databases experiments show the effectiveness of this algorithm.