Font Size: a A A

A Nonlinear Canonical Correlation Analysis Based On Extreme Learning Machine With Applications

Posted on:2019-05-29Degree:MasterType:Thesis
Country:ChinaCandidate:X H WenFull Text:PDF
GTID:2348330569979542Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
With the ever-growing development of sensor technology,the way of data acquisition is more diverse.Data obtained through sensors usually with different modalities,such as images,text,and sound signals.How to get people's interest from these different forms of information and achieve synergies and utilization of the information has become an imminent problem.Canonical Correlation Analysis(CCA)is a powerful tool for information fusion.It mainly focuses on the correlation problem between two sets of variables.The complex correlation between two sets of variables can be simplified through CCA,and these few simplified pairs of variables will be used to reflect the mutual information of original variables.However,CCA has a great limitation that only implements the linear transformation of data.Kernel CCA can learn the nonlinear relationship between variables using kernel method,but the kernel function is selected manually.And the large scale of the kernel matrix may cause computational difficulties.It is undeniable that neural networks have achieved great success as a nonlinear mapping system.But the traditional gradient-based algorithm has a slow training process and poor generalization.Extreme learning machine(ELM)is a single hidden-layer feed-forward neural network,with the advantages of simple structure,fast learning speed and good generalization ability.This paper introduces ELM into CCA,maximizing the correlation between the nonlinear features of different modality data learned by multilayer ELM neural network.Therefore,the nonlinear problem is indirectly transforming into a linear problem.As a result,not only does the training speed have significant improvement,the fusion features for object recognition tasks can also get satisfactory performance.The methods based CCA require restrictive assumptions,such as a priori known pairings between all data samples.However,in practice,it is inevitable that there will be some missing samples in a certain modality or disorder of samples between different modalities.Generally,the data that only have category level correspondences rather specific sample correspondences are called weakly paired multimodal data.In view of the above problems,this paper will solve the problem of feature fusion in weakly paired cases using a weakly paired maximum covariance analysis method based on multilayer ELM.The main research work of this paper includes the following aspects:(1)A new nonlinear canonical correlation analysis method is proposed.The ELM is introduced into the CCA framework to learn the complex nonlinear representations between two sets of variables.Compared with the existing nonlinear CCA method based on deep neural network,the proposed method not only significantly improves the training speed,but also effectively improves the correlation between features.(2)Aiming at the weakly paired cases in multi-modal data,a weakly paired multimodal fusion framework is proposed.Potential feature representations are extracted for each modality by multilayer ELM.In order to obtain the projected vectors of the weakly paired features in the common subspace,the pairing matrix is introduced into the maximum covariance analysis.Therefore,the weakly paired problem of modality samples is effectively solved,and the obtained projected vectors have a strong nonlinear correlation.(3)Experimental verification and analysis are performed on the current public large-scale multimodal datasets(two RGB-D datasets and a real-world weakly paired scene classification dataset).
Keywords/Search Tags:Canonical correlation analysis, Extreme learning machine, Multi-modal fusion, Weakly paired data, Feature extraction, Object recognition
PDF Full Text Request
Related items