Research On Multi Manifold Learning Methods Under Uncorrelated Constraints

So far,dimensionality reduction remains a research hotspot in the field of data mining,pattern recognition,and image processing.Manifold learning,as the main representative method of nonlinear dimensionality reduction,performing well in dealing with nonlinear data sets,but still contains some defects.For example,most manifold learning algorithms discriminant features with high correlation,high redundancy and insufficient learning performance in the features extraction,which is not conducive to subsequent recognition.In order to solve this problem,this paper puts forward two kinds of manifold learning algorithm based on uncorrelated constraint,both can effectively reduce feature redundancyIn this thesis,the main work includes the following two aspects:(1)Propose a multi-manifold learning algorithm based on global uncorrelated.In order to improve the learning ability of the algorithm,the distance measurement from feature space to feature space is introduced.Firstly,calculate the projection of all data points in the heterogeneous nearest neighbor feature space.At the same time,propose the global uncorrelated constraint criterion to reduce the redundancy of discriminating information and obtain the optimal projection matrix by maximizing the spatial distance of heterogeneous features under the global uncorrelated constraint.(2)Propose a subspace learning algorithm base on locally uncorrelated.In order to mine the local discriminant information of sample points,propose a local uncorrelated constraint criterion based on the distance from the point to the feature space.For each sample points,determine the projection of the nearest-neighbor feature space within the class and the nearest-neighbor feature space between the class,then obtain the distance matrix from the point to the feature space and reuse it to build local scatter within class and scatter between classes.At the same time,deduce the local margin criterion and local uncorrelated constraint based on the distance matrix of point to feature space,then explore a local uncorrelated discriminant subspace by maximizing the local margin under the local uncorrelated constraint.The results of several commonly used face data sets show that the multi-manifold learning algorithm proposed in this paper is better than other advanced algorithms in face recognition,and can reduce the redundancy of feature information and improve the ability of distinguishing features.