Study On Multiset Canonical Sparse Cross-view Correlation Analysis

As a multi-modal feature extraction method,canonical correlation analysis is receiving more and more attention.It extracts optimal projection features by maximizing the relevance of the two sets of feature samples.We introduce sparse representation into CCA algorithm and based on MCCA algorithm.Details of the work are as follows:(1)As sparse representation can reduce computational complexity and saving storage space,we select local information adaptively,by using sparsity preserving,and class information is embedded into the algorithm through calculating the weight matrix in samples of the same class,which makes it a supervised method.Besides,we introduce the idea of cross correlation which enhances the relationship between unpaired samples of different views.We proposed canonical sparse cross-view correlation analysis with missing samples(CSCCAM)algorithm.The experiments on multi-character handwriting data set,CENPARMI and PIE face database to verify the effectiveness of the proposed algorithm.Experiments show that our algorithm robust to the number of missing samples.(2)We embarks from the fusion of multi-group feature which is based on CCA algorithm,combine CSCCA algorithm,propose a multi-set canonical sparse cross-view correlation analysis with missing samples(CSMCCAM).The experiments on multicharacter handwriting data set and CENPARMI database results show that the recognition accuracy of CSMCCAM is relatively insensitive to the number of missing samples.(3)Traditional CCA suffers the small sample size(SSS)problem due to high matrix dimension from the matrix-vector preprocessing.This problem was solved by a new supervised learning method called 2D-DMCCA.The 2D-DMCCA computes image covarianced matrix directly without matrix-to-vector conversion.This method effectively solves the SSS problem.The experimental results on AR and ORL face database show that the 2D-DMCCA algorithm is effectively than others.