Research On Manifold Learning Feature Extraction Method Based On Subspace And Its Applications To Face Recognition

The feature dimensions of the original face image are usually high,and it contains a lot of irrelevant redundant information.How to effectively reduce the dimension of face image and extract the key features from face image,which is the critical step in the face recognition process.Subspace method based on manifold learning is a mainstream method in face recognition.The idea of this method that face data from high dimensional space can be expressed in low dimensional manifold space aiming at preserving internal structure and regularity of the face data.Starting from the idea and principle of manifold learning,based on principle and related knowledge of manifold learning algorithms,the paper explored new algorithms which have more excellent performance aiming at the deficiency and defect existing in face recognition algorithms.The main contents and results of the paper are as follows:Acording to the traditional composition method of k nearest neighbor which has the problem of choosing the parameter and compressing the picture into the one-dimensional form that may ignore the original structure of the samples.On the basis of the theory of LPP,adopting the composition method of adaptive neighborhood graph and introducing the column structure information,this paper proposed a locality preserving projection algorithm based on column information of sample and adaptive neighborhood graph.In this method,corresponding column neighbors of column samples are determined adaptively by the column information of the samples,and then the neighbors of the sample are determined adaptively by the number of coupled column neighbors between two samples.This algorithmic can avoid the difficulty of choosing nearest neighbor parameter k by using the original structure information of sample,finally truly show themanifold structure of sample.According to the difficulty of reconstructing the face image in High-dimension,nonlinear structure,through combining nuclear technology and LPP algorithm meanwhile introducing the theory of orthogonalization,the paper proposed kernel orthogonal global discrimination and locality preserving projection(KOGDLPP)based on difference form.In this algorithmic,the outer-class similarity matrix reflects the global structure,which was described by the distance between the sample and center point of different classes;the inner-class similarity matrix reflects the local structure of sample space,which was described by the distance between the sample and center point of same classes.Finally,by introducing the idea of orthogonalization,solve reconstructed objective function to obtain orthogonal projection vectors which can remove redundant information and enhance the ability to reconstruct for data.This paper studied the small sample problem which existing in the local maximal margin discriminant embedding algorithm(LMMDE),with introducing the theory of multi-manifold and cosine angular distance,the paper proposed multi-manifold Maximal Margin Discriminant Preserving Embedding based on Cosine Angular Distance(CMMMMDPE).Firstly,we divide the sample into several local samples,establish manifold for every sample and construct multi-manifold in the whole sample space.Then reduce the dimensionality of manifold sample in the multi manifold space to find the optimal matrix which is simplified to improve the recognition rate of the algorithm in small sample.Make full use of class information,replace euclidean distance in traditional method with the cosine distance during determining multi-manifold outer-class neighborhood and multi-manifold inner-class neighborhood for the small sample,which can make the algorithm to be insensitive to the neighbor parameter k and improve algorithm's feasibility and effectiveness for outlier sample points.