Data Dimension Reduction And Dictionary Learning And Its Application In Human Feature Recognition

Human feature recognition based on video image data(such as face recognition,behavior recognition,and so on)is one of the research hotspots in the field of computer vision,the application requirements in many fields have promoted the development of this technology,such as video surveillance,human-computer interaction,video retrieval,sports analysis,and medical diagnosis etc..However,at present,most video image data collected by cameras are high-dimensional data,and it is difficult to obtain effective behavior features directly.Moreovre,individual differences and external environment changes lead to a large number of variables,and the existence of these variables makes it difficult to accurately describe the human's appearance and motion features.Therefore,through developing effective image data dimension reduction and sparse representation as well as machine learning methods,to improve the accuracy and robustness of human body feature extraction and recognition has become the key issues.This thesis mainly focuses on high dimensional video image data to carried out theoretical and experimental research,which contains nonlinear dimensionality reduction methods based on manifold learning,dictionary learning methods based on compressed sensing and sparse representation,and its application in face recognition,human behavior feature extraction and behavior pattern classification.Firstly,several classical nonlinear data dimensionality reduction methods based on manifold learning are analyzed.Meanwhile,the dimensionality reduction effect evaluation methods are analyzed.To solve the problems of using subjective analysis method to evaluate the dimensionality reduction effects,including strong subjectivity and lack of necessary quantitative calculation for guidance,a dimensionality reduction effect evaluation method is proposed.Two indicators are utilized to quantify the trustworthiness of visual effect diagram and the neighborhood retention property,they are trustworthiness and continuity.Based on them,the quantitative evaluation for dimensionality reduction effect is realized,meanwhile the validity of the proposed methods is verified by the experiments on several classical datasets,such as Swissroll and Helix,and so on.Secondly,to deal with the incremental dimensionality reduction problem of large data sets,a method of nonlinear dimensionality reduction and reconstruction based on locality constraint dictionary learning is proposed.This method utilizes local neighborhood constrained conditions to construct a non-convex objective function.Its analytical solutions can be obtained by the Lagrange multiplier method,and alternate iteration method is used to optimize the objective function,which can reduce the computation time and the storage space,and the efficiency of optimization objective function is improved.Experimental results on the classical datasets show the effectiveness of the proposed method,such as Swissroll and Swisshole.Furthermore,an image classification method based on locality constraint dictionary learning is proposed in this paper.By constructing a category based structured dictionary,sparse representation coefficients and reconstruction errors of the testing sample can be used for classification.Moreover,its locality constrained conditions can enforce the intra-class distance,which can improve the discriminative ability of the structured dictionary,and then the recognition performance is improved.In addition,the experimental results of face recognition on ORL,Extended YaleB and AR datasets,gender classification on AR gender dataset,and object category classification on Caltech101,Caltech256 and Cifar-10 datasets all demonstrate the validity of proposed method.Thirdly,this paper aims at the human behaviors' high complexity and variability,a method based on energy image species and pyramid histogram of orientated gradients(PHOG)feature is developed.The proposed method calculates the averaged motion energy image(AMEI)and enhanced motion energy image(EMEI)of the target contour image firstly,and extracts their PHOG features as the description of human behavior characteristics.After that,support vector machine(SVM)and Real Adaboost algorithms are used to design multi-class classifiers for human behavior recognition,respectively.The Real Adaboost algorithm based on look-up-table type weak classifier has good anti-overfitting ability and can improve the performance of human behavior recognition.The experimental results verify the effectiveness of the proposed method.Finally,to describe the human behavior effectively and improve the accuracy of recognition,human behavior recognition methods based on two-dimensional spatial template are studied,because of the two-dimensional spatial-template is global description of human behavior,which makes it difficult to express local information accurately,a human behavior recognition method based on local constrained linear encoding is proposed.In this method,we first segement the two-dimensional spatial template by spatial pyramid matching(SPM),and extracts the SIFT feature of all patches,then a multi-layer patches descriptor is constructed for human behavior description.And then,encodes them by locality constrained linear encoding and recognize the behaviors by Max pooling.In addition,a human behavior method based on locality constrained dictionary is proposed,which uses the improved energy image species to describe the behavior characteristics,and constructs a discriminative structure dictionary for recognition.The two methods are experimental studied on Weizmann and DHA datasets respectively,and the effectiveness of the methods are verified.