Research On Face Recognition Based On Multi-manifold Learning

With the birth of the computer in the last century,computer technology has been developed rapidly,and now enters the era of information explosion.In daily life and work,more and more people come into contact with massive complex high-dimensional data,which are mainly characterized by large data volume,high dimension,high data growth rate and nonlinear distribution.The storage required for processing these data is large and the running speed is slow,especially in image recognition and artificial intelligence.As a result,a "Dimensional disaster" was triggered.How to find the essential characteristics of these high-dimensional data and then analyze and reduce the dimensionality of high-dimensional data has become an urgent problem to be solved.Feature extraction is an effective data processing tool.The purpose of dimensional reduction is to reduce the dimension of the data,embed the high-dimensional data into the low-dimensional subspace,and hope to retain the essential information and internal structure of the data.It has been widely used to solve dimension disaster problems in data dimension reduction problems,but most of the existing manifold learning is effective for single-sampled data,but has certain limitations for data sampled from multi-manifold.Therefore,in recent years,some scholars put forward the multi-manifold learning method.The multi-manifold learning method embeds high-dimensional data into the corresponding low-dimensional subspace,and hopes to retain the internal structure and essential characteristics of the data,discover the essential characteristics of high-dimensional data hiding,and achieve the purpose of classification and clustering.Multi-manifold learning has been widely used to discover the internal structure of data,and in recent years,it is mostly used for image recognition and analysis.For labeled datasets,this paper proposed the supervised linear multi-manifold learning method,using sample label information to increase the degree of aggregation samples within a manifold,reduce the degree of coupling between samples,reducing the overlap between the manifold,improving the maximum recognition rate.The traditional multi-manifold learning method uses Euclidean distance measure degree of neighbor points between inter-manifold.However,manifolds are local Euclidean spaces.Based on this,this paper proposes to use Euclidean distance to measure the similarity degree of data points in intra-manifold,and obtain the similar nearest neighbor sets on the intra-manifold,uses the method of graph matching to measure the similarity degree of data points in inter-manifolds,to obtain the set of similar nearest neighbors of data points on the inter-manifold.Establish an optimization function to make the distance between neighboring points on the inter-manifold in the low-dimensional representation larger and larger,and the weighted distance between neighboring points on the intra-manifold getting smaller and smaller.The data points on the intra-manifold compactly surround the center of the manifold.The method proposed is compared with the three traditional feature extraction methods.The results show that the algorithm has a high average recognition rate,which has certain advantages and effectiveness.For insufficiently labeled datasets,this paper proposed the semi-supervised multi-manifold learning method to solve the problem of lack of labeled samples in the case of multi-manifolds.First,a subgraph division method based on label propagation is proposed.This method obtains the index set,and divides the index set to obtain multiple subgraphs,and assigns labels to unlabeled samples.Based on LDA,the local divergence matrix and the intra-manifold divergence matrix are established to ensure that the neighboring points on inter-manifold are more dispersed and the data points on intra-manifold are more compact.The method proposed in this paper is compared with three semi-supervised feature extraction methods.Experimental results show that the method has better advantages in terms of clustering accuracy and data visualization.