Structure Embedding Extended Correlation Projection With Applications To Image Recognition

By ext.ractirng available inf'ormation or experience from finite observational samples with the help of the data analysis,pattern recognition(PR)models the problem at hand and deals with different patterns automatically.Thus,feature extraction is one of the most basic problems of PR.With more and more different data processing and description skills,the same objects always have multiple representations from different feature spaces(views or modalities).Since conventional feature extraction or dimensionality reduction methods,e.g.,principle component analysis(PCA)and linear discriminant analysis etc.,mainly focus on dealing with single representation data,they are not suitable to applied to reflect different characteristics of the object.Therefore,in this dissertation,we mainly focus on joint correlated projection of multi-representation data based on canonical corre-lated analysis(CCA),expecting that extracted features can not only obtain the effectively discriminative information,but also eliminate the redundant information to a certain ex-tent in each feature representation.Utilizing semi-supervised learning,fractional-order embedding and collaborative representation,etc.,we extend related theories into joint feature extraction in order to enhance the discriminative ability of extracted features.Our work mainly includes the following parts:(1)Single sample per person reeognition is one of the most challenging problems in face recogition due to the lack of ihnformation to predict the variations in the query sample.To address this problem,we present in this dissertation a novel face recogni-tion algorithm based on a robust collaborative representation and probabilistic:graph model,which is called Collaborative Probabilistic Labels(CPL).First,by utilizing la-bel propagation,we construct probabilistic labels for the samples in the generic training set corresponding to those in the gallery set,thus the discriminative information of the unlabeled data can be effectively explored in our method.Then,the adaptive variation type for a given test sample is automatically estimated.Finally,wc propose a novel.reconstruction-based classifier for the test sample with its corresponding adaptive dictio-nary and probabilistic labels.The proposed probabilistic graph based model is adaptivcly robust to various variations in face images including illumination,expression,occlnsion,posc,etc.,and is ablo to rceduce rcquircd training images to one sample per class Exper-imcntal results on scvcral widely used face databases demonstrate that,companrcd with traditional methods,CPL is more diseriminatve and robust.(2)When facing semi-supervised multi-modal data which widely exist in real-world applications,CCA usually performs poorly due to ignoring useful supervised informa-tion.Meanwhile,due to the limited labeled training samples in the semi-supervised scenario,supervised extensions of CCA suffer from overfitting.In this dissertation,we present a robust multi-modal semi-supervised feature extraction and fusion framework,termed as dual structural consistency based multi-modal correlation propagation proje.c-tions(SCMCPP).SCMCPP guarantees the consistency between representation structure and hypotaxis structure in each modality and ensures the consistency of hypotaxis struc-ture between two different modalities.By iteratively propagating labels and learning affinities,discriminative information of both given labels and estimated labels is utilized to improve the affinity construction and infer the remaining unknown labels.Moreover,probabilistic within-class scatter matrices in each modality and probabilistic correlation matrix between two modalities are constructed to enhance the discriminative power of features.Extensive experiments on several benchmark face databases demonstrate the effectiveness of our approach.(3)Due to the noise disturbance and limited number of training samples,within-set and between-set sample covariance matrices in CCA based methods usually deviate from the true ones.This dissertation re-estimates the covariance matrices by embed-ding fractional order and incorporate the class label information.First,we illustrate the effectiveness of the fractional-order embedding model through theory analysis and ex-periments.Then,we quote fractional-order within-set and between-set scatter matrices and incorporate the supervised information,presenting novel fractional-order embedding generalized canonical correlations analysis(FEGCCA)and fractional-order embedding discriminative canonical correlations analysis(FEDCCA).Extensive experimental result-s on various handwritten numeral,face and object recognition problems show that our methods.ds are very effective and robust to small sample size(SSS)problems in terms of classification accuracy.(4)CCA always fails to work well when facing SSS problems,because the noise dis-turbance and limited number of training samples usually contribute to lacking information of within-class variations.To acddress this problem,in this paper,we first reveal the re.-construction relationship between overall and within-class covariance matrices.Then,an Adaptive Compensation Correlation Projection(ACCP)method,which constructs adap-tive within-set and between-set scatter matrices by adapting generic discriminant model,is proposed.As a specific implernentation of the ACCP,a novel Probabilistic Varia-Lion Subspace(PVS)model is also proposed to infer,based on the generic traiuing,set,the adaptive subspace for representing a specific gallery sample.Extensive experimental results on five benchmark face and object databases show that ACCP is very effective and outperforms existing joint dimensionality reduction or feature extraction methods in terms of classification accuracy.(5)As a generalized extension of CCA,multiset CCA(MCCA)can bandle data that is represented by more than two views.However,most existing MCCA-related meth-ods fail to discover the intrinsic discriminating structure among data spaces.To deal with this problem,by taking discriminative information of within-class and between-class sparse reconstruction into account,we present sparse discrimination based multiset canonical correlation(SDbMCC)model.It has been proved that it is the collaborative representation but not the l1-norm sparsity that makes sparse representation based clas-sification powerful.Furthermore,considering the correspondence between multiple views and unsupervised scenario,we present view-consistent collaborative preserve projection(C2PP)mechanism along with its supervised extension,based on which we construct view-consistent collaborative multiset correlation projection(C2MCP)and view-consistent col-laborative discriminative multiset correlation projection(C2DMCP)frameworks.In ad-dition,we discuss in depth the relationships between C2DMCP and other multi-view subspace learning methods,i.e.,MCCA and SDbMCC.Extensive experimental results in pattern recognition tasks demonstrate the effectiveness of the proposed methods.