Font Size: a A A

Study On Multi-view Learning In Visual Identity

Posted on:2016-10-04Degree:MasterType:Thesis
Country:ChinaCandidate:H L LiuFull Text:PDF
GTID:2348330536954741Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of data collection and storage technology,a large number of videos or photos are accessible,and then the research of visual identity has become the focus of attention.The methods of visual identity deal with massive and complex information,at the same time,study new theory and advanced technology,so that the instructions of target recognition can be completed more intelligently.Now they have been widely used in industry,public security,commerce,medical,military and agriculture.Multi-view learning can make the most of multiple features and more fully reflect the nature of the targets.Different views can complement each other and improve the learning performance,so multi-view learning methods can stretch the limitations of the traditional learning methods based on single feature.Therefore,multi-view learning provides a new way for problems solving in visual identity.Based on the in-deep research on multi-view learning,the main contributions of this thesis are as follows:1.Multi-view regularized logistic regression method is proposed.Compared with the traditional support vector machine(SVM)method,the function of logistic regression is smooth,sensitive to abnormal points and suitable for huge amount of data and so on.Therefore,we propose a multi-view regularized logistic regression model,in which different features are regarded as different views,and obtain some satisfying results in action recognition.2.Multi-view sparse coding method is proposed.Considering the different contributions of different facial components to face analysis,a face analysis based on multi-view sparse coding is proposed,in which facial components are regarded as different views.On the other hand,Gabor wavelet features that displace facial gray features are used to improve the recognition rate by combining multi-view sparse coding.3.Laplacian-Hessian regularization method is proposed.In the paper,we propose Laplacian-Hessian regularization method by analyzing the different features between Laplacian and Hessian regression in image classification of semi-supervised learning,which can improve performance by making the best of the superiorities of Laplacian and Hessian.
Keywords/Search Tags:multi-view, semi-supervised learning, manifold learning, logistic, sparse coding, Hessian
PDF Full Text Request
Related items