Research And Application For Feature Extraction On Discriminative Two-Dimensional Canonical Correlation

In recent years,data representation has become more and more diverse.Especially in the filed of pattern recognition,where the same object can be characterized by various feature vectors in different high-dimensional feature spaces.Canonical correlation analysis(CCA),as an unsupervised feature learning algorithm,is required to convert the two-dimensional matrix representation of the image into a one-dimensional vector form during feature extraction process.This transformation will lose information about the intrinsic spatial structure of the pixels and increase the complexity of the computation.In addition,the CCA does not consider the class label information of the training samples,making it difficult to extract discriminative low-dimensional features.In view of this,in this paper,the discriminative extension of two-dimensional CCA is studied in depth,and a discriminative two-dimensional canonical correlation feature extraction algorithm is established,using multi-view data as the object of study,combined with the ideas of supervised learning,fuzzy set theory,and multiple sets,and applied to image recognition.The main research results and innovations of this paper are as follows:(1)A supervised two-dimensional canonical correlation feature extraction algorithm(Supervised Two-Dimensional CCA,S2CCA)is proposed.The algorithm takes into account both inter-view and intra-view sample class label information to maximize the correlation of canonical projections between similar samples.The S2CCA algorithm has stronger discriminative power and the recognition rate has been improved to a certain extent when compared with the original two-dimensional CCA algorithm on the face datasets of AR,AT&T and CMU PIE.(2)A generalized fuzzy two-dimensional canonical correlation feature extraction algorithm(Generalized Fuzzy Two-dimensional CCA,GF2DCCA)is proposed.This algorithm uses fuzzy K-nearest neighbors(FKNN)to generate a membership matrix that stores the class membership values of all sample points.And on this basis,the fuzzy intra-class scatter matrix and inter-group covariance matrices are redefined,and fuzzy sample coding information is embedded into a two-dimensional canonical correlation model for learning.GF2DCCA takes into account the two-dimensional spatial structure information of the image matrix while considering the degree of affiliation of actual observations with different classes According to the experimental results,GF2DCCA is able to make the complexity of the calculation less while minimizing the loss of image spatial structure information,especially in low-dimensional space,to obtain a two-dimensional representation with strong discriminative power.(3)A double-directional multiset canonical correlation analysis(Double-Directional Multiset CCA,D2MCCA)is proposed.On the basis of which,drawing on the idea of generalized canonical projection idea,this paper further proposes a discriminative double-directional multiset canonical correlation feature extraction algorithm(Discriminative D2MCCA,D-D2MCCA).D2MCCA extends the correlation between two views to the correlation between several view samples,considering the row and column spatial information of two-dimensional images;while D-D2MCCA introduces the class label information of the training samples,minimizing the intra-class scatter matrix while maximizing the total correlation of multiple set.The experiments imply that the recognition rate is much more better compared to D2MCCA.