The Canonical Correlation Analysis Algorithm Based On Sparse Representation

As a multi-modal feature extraction method, canonical correlation analysis (CCA) is increasingly receiving more and more people’s attention. It extracts optimal projection features by maximizing the relevance of the two sets of feature samples. Many excellent classification and recognition algorithms are derived from this. People enriched algorithm theories based on CCA for pattern recognition by introducing some skills such as kernel skills、locality preserving、sub-mode、supervised information and so on. As compression perception theory was put forward, the concept of sparsity representation have been introduced into the field of pattern recognition in recent years. Many researchers joined them, and proposed some excellent algorithms based on sparsity representation theory. This article is to make a deep study of the related algorithms combining sparsity representation ideas with algorithm theory about CCA. Details of the work are as follows:(1)Beginning from traditional sparsity preserving projection (SPP)、sparsity preserving canonical correlation analysis (SPCCA), we find the lack that the supervised informations introduced into them are inadequate. Learning the method of introducing supervised informations in supervised sparsity preserving projection (S PP), we make improvements and propose supervised sparsity preserving canonical correlation analysis (S2PCCA).This method introduce supervised information into SPCCA and achieve good recognition effect.(2)We look at the sparsity representation methods of introducing supervised information from another angle. Beginning with existing sparsity representation discriminative analysis (SRDA), we propose generalized sparsity discriminative canonical correlation analysis (GSDCCA). This algorithm introduces supervised information into SPCCA by using another different idea and also achieves good recognition effect.(3)Locality discriminative canonical correlation analysis (LDCCA) is a proposed method which combines their respective advantages of locality preserving canonical correlation analysis (LPCCA) and discriminant canonical correlation analysis (DCCA). This article analyzes the shortcomings of the method’s introducing class informations. And learning from two methods of introducing supervised informations above mentioned, we propose two new improved methods and improve the recognition rates.(4) We introduce sparsity representation ideas and supervised informations into multiset canonical correlation analysis (MCCA). We propose supervised sparsity preserving multiset canonical correlations analysis (S2PMCCA) and sparsity discriminative multiset canonical correlations analysis (SDMCCA). The experiments verified the effectiveness of the methods.