Image Set Representation And Classification With Graph Model

With the rapid development of computer software and hardware,a large amount of image or video data have been generated.How to effectively and quickly analyze those data becomes a key research issue in computer vision and pattern recognition area.In recent years,image set object recognition based problem has attracted more and more interesting.Compared with single image based recognition problems,image set can provide more effective and overall information because an image set contains numerous image samples.Image set based recognition problem has two key challenges:(1).How to represent an image set;(2).How to classify the image sets.In computer vision and pattern recognition area,graph theory has very successful applications.Graph model can effectively describe an object,including the features of object or the relations between objects.However,the traditional graph theory is usually used in single image based problems and cannot be applied to solve the image set based recognition problem.This thesis applies graph theory to image set based recognition problems and explores image set representation and classification methods based on graph model.The main contents are introduced as follows.(1)For image set representation,an innovation graph method is proposed to represent an image set,named Covariate-Relation Graph(CRG).The node of CRG represents the covariate of image set data matrix and the edge reresents the relation(such as similarity)of covariates.The relation of covariates can be measured with Linear function,Laplacian function or Gaussian function.In terms of image set classification,Linear Discriminant Analysis(LDA)is introduced to learn the classification or recognition tasks.Before using LDA to do the classification task,the similarity between image sets is measured.As we all know,the numerous data is nonlinear which may lead to computationally expensive.The kernel method can map the data from Riemann manifold to European space and thus solve the nonlinear problem.Then we can use traditional metrics to learn the similarity measurement in Euclidean.In experiment,some version experiments are evaluated to demonstrate the effectiveness and robustness of our methods on four commonly used datasets.(2)In terms of image set representation,a new graph model is proposed to represent an image set based on CRG model,and it is named Attributed Covariate-Relation Graph(ACRG).Based on ACRS Graph Sparse Representation Classification(GSRC)is proposed to learn the image set classification problem.ACRG both considers the feature of covariate and the relation of covariates,i.e.,nodes and edges of the graph,respectively.The first-order(node)and second-order(edge)information are useful to learn the recognition problem.After representing the image set with ACRG,a new method named-Graph-Sparse Representation Classification(GSRC)is proposed to learn the image set classification.The Attributed Covariate-Relation Graph(ACRG)is reconstructed with Graph Sparse Representation(GSR)model.Then the solution and the optimal process of the GraphSparse Representation(GSR)algorithm are introduced.Finally this thesis gives the proof of convergence and the classification criterion.In terms of experiment,it demonstrates the effectiveness-and robustness of ACRG with four adding noisy datasets.(3)For image set representation,this thesis proposes a method to represent image set based on Low-Rank Subspace and Covariate-relation Graph(LRSCRG)model.The number of images in an image set is very large and the images are different from each other,and we should use different subspace to represent the image set.To solve above problem,we firstly use low-rank method to represent an image set,and then use spectral clustering method to obtain the subspaces.Each subspace is represented with Covariate-relation(CRG)model.In terms of image set classification,Linear Discriminant Analysis is used to learn the classification tasks,and a measurement for the similarity of image sets based on Low-Rank Subspace and Covariate-Relation Graph representation model is introduced.In experiment,the results show the effectiveness and robustness of our LRSCRG method.