| Image feature extraction is one of the main research problems of pattern recognition.For the recognition of high-dimensional noise image,the recognition effect depends heavily on the effectiveness of the extracted features.Linear discriminant analysis and principal component analysis are the two most representative feature extraction methods,which have been widely used in image recognition.However,these two methods mainly deal with one set of features,and cannot extract features from multiple sets of data at the same time.As a feature extraction method based on two sets of features,canonical correlation analysis(CCA)can extract useful features from images by using the maximum correlation between two sets of features.The main work is as follows:(1)l0-based two-dimensional sparse preserving canonical correlation analysis(2D-SPCCA/l0)and l0-based two-dimensional sparse locality preserving canonical correlation analysis(2D-SLPCCA/l0)are proposed.Two-dimensional sparse preserving canonical correlation analysis can preserve the sparse reconstruction between samples in the process of dimension reduction,but the sparse coefficient based on l1-norm solution can not produce sparse results.Inspired by this,2D-SPCCA/l0 and 2D-SLPCCA/l0 are proposed,which not only realizes the effective fusion of two groups of features,but also constrains the sparse reconstruction between features,and improves the ability of feature identification.On the basis of 2D-SPCCA/l0,manifold learning structure is introduced,and 2D-SLPCCA/l0 is proposed.Experiments on face data sets show that the two methods have good discrimination performance.(2)l0-based two-dimensional supervised sparse preserving canonical correlation analysis(2D-SSPCCA/l0)and l0-based two-dimensional supervised sparse locality preserving canonical correlation analysis(2D-SSLPCCA/l0)are proposed.Both 2D-SPCCA/l0 and 2D-SLPCCA/l0 are unsupervised algorithms,which don’t use the label information of samples and waste a lot of discrimination information.Therefore,we introduce label information into 2D-SPCCA/l0 and 2D-SLPCCA/l0,and propose 2D-SSPCCA/l0 and 2D-SSLPCCA/l0.Experiments on face data sets show that these two methods can effectively extract discriminative image features(3)l0-based two-directional two-dimensional sparse preserving canonical correlation analysis(2D2-SPCCA/l0)and l0-based two-directional two-dimensional sparse locality preserving canonical correlation analysis(2D2-SLPCCA/l0)are proposed.2D-SPCCA/l0 and 2D-SLPCCA/l0 can only deal with matrix sample sets with the same number of rows or columns,but can not deal with matrix sample sets with different sizes.To overcome this problem,2D2-SPCCA/l0 and 2D2-SLPCCA/l0 are proposed.Detailed experiments on face data show the effectiveness of the proposed method.(4)l0-based two-directional two-dimensional supervised sparse preserving canonical correlation analysis(2D2-SSPCCA/l0)and l0-based two-directional two-dimensional supervised sparse locality preserving canonical correlation analysis(2D2-SSLPCCA/l0)are proposed.Inspired by 2D-SSPCCA/l0 and 2D-SSLPCCA/l0,2D2-SSPCCA/l0 and 2D2-SSLPCCA/l0 are proposed by adding tag information to 2D2-SPCCA/l0 and 2D2-SLPCCA/l0.Experiments on face data sets show that 2D2-SSPCCA/l0 and 2D2-SSLPCCA/l0 can extract more discriminative image features. |
