Research On Image Representation Based On Correlation Projection Analysis

Image representation is one of the important foundations of computer vision,pattern recognition,and machine learning.The linear methods of image representation have been conducted in-depth study by domestic and foreign experts.With the improvement of image acquisition technology,more and more features can be collected,how to deal with a variety of features to represent the image has become one of the hot spots.The correlative projection analysis is a traditional multi-view linear subspace learning method,which aims to find two sets of projection directions,so that the correlation of the features in the correlative space is maximized.Therefore,the correlative projection analysis naturally become a multi-view learning method.In recent years,the study of correlative projection analysis has received extensive attention.In this paper,we focus on the framework of canonical correlation analysis(CCA)which is one method of correlative projection analysis,and the related improvements that how to extract and fuse more effective features from image are putted forward,and further,the problem that CCA has the weak constraint on projection vectors is discussed and the related researches are studied.1.The problem of robustness of CCA based on multiple linear regression is studied.According to the problem that the target optimization function of traditional CCA based on L2 norm is not robust to the outliers of the image feature.On the basis of defining the relevant error,the generalized mean is introduced into the equivalent form of CCA based on multiple linear regression,and the target optimization function is reconstructed.Robust canonical correlation analysis based on generalized mean(GMCCA)is proposed.And GMCCA obtains the projection vectors by iterative method,that avoids solving the inverse of the sample covariance matrix,thus GMCCA solves the "small sample problem".The experimental results on the real databases(MFD,ORL and COIL20)show that the algorithm can effectively suppress the influence of the outliers and extract the robust canonical correlation features.2.Weighted correlation analysis with feature information(WCCA)is proposed.Inspired by the local feature selection proposed by Narges et al,the variance information of the feature and the information of canonical correlation after extraction are fully used to make up for the problem that CCA can not distinguish the main and secondary features.On the one hand,the original feature is weighted by the variance information,and on the other hand,the extracted feature is weighted by canonical correlation.To reduce the impact of external factors on the features,the choice functions are taken.The experimental results show that WCCA can effectively use the attributes of the feature to extract more discriminant features and improve the performance of the algorithm.3.Coherent-projected canonical correlation analysis(CPCCA)based on the uniform size of the projection vectors is proposed for the weak constraint of the traditional CCA to the projection vectors.On the basis of relaxing the weak constraint condition of the original CCA to the projection vectors,we constrain the uniform size of the corresponding projection vectors.In the process of solving the parameters,the relationship and difference betweenCPCCA and CCA are discussed,and the physical meaning of the parameters is revealed.A two-steps approach for parameters is proposed.The experimental results of multi-feature handwritten database(MFD),Japanese female facial expression database(JAFFE)and face database(YALEB)verify the effectiveness and feasibility of the algorithm.