Feature Extracion And Application Based On Canonical Correlation Analysis

Currently,image recognition has become one of the most important and active research topics in the field of artificial intelligence and pattern recognition.This paper is based on canonical correlation analysis.In the framework of canonical correlation analysis,multi-group features,several kernel functions and sparsity preserving projection are introduced,thus canonical correlation analysis is extended to nonlinear space and classification information is introduced to improve classification performance.The research results are applied to face recognition and handwritten numeral recognition in this paper.Mainly,it includes the following aspects:1.Feature fusion method based on multi-set kernel canonical correlation analysis.Canonical correlation analysis is to study the linear relationship between two groups of random variables.However,due to the complexity of the real world,it is impossible to solve the problem with a few simple linear functions.Therefore,aiming at the difficulty of extracting enough features and complex nonlinear information by canonical correlation analysis,we propose the multi-set kernel canonical correlation analysis.Multi-group features and kernel function are introduced to extending two sets of features to multi-group features,and mapping image data to high-dimensional feature space in order to extract the more discriminant information and nonlinear information and get more discriminated features.The experimental results on a multi-face databases and handwritten numeral datasets show the effectiveness and stability of the proposed method.2.Pattern recognition based on sparsity preserving canonical correlation analysis.Canonical correlation analysis is a kind of unsupervised methods,label information of samples is not used,so it wastes a lot of useful information and has poor generalization ability.To solve this problem,we propose the sparsity preserving canonical correlation analysis.In the framework of canonical correlation analysis,the class information of samples is introduced by using sparsity preserving projection,and the correlation characteristics of two sets of feature vectors are classified as effective information,which improves the recognition accuracy.The experimental results show that the SPCCA algorithm achieves better recognition performance compared with other related algorithms.3.Image recognition based on kernel sparsity preserving canonical correlation analysis.The kernel sparsity preserving canonical correlation analysis algorithm we proposed is a combination and optimization of the KCCA and SPCCA algorithm.Firstly,image data samples are mapped to high dimensional feature space by kernel function.Secondly,feature extraction and fusion in the kernel space is studied by using the sparsity preserving canonical correlation analysis algorithm.Thirdly,the obtained features are used for classification and recognition.This method takes into account the nonlinear information in data samples and class label information of samples.The experimental results obtained on a number of public face databases and handwritten numeral datasets demonstrate the effectiveness of the kernel sparsity preserving canonical correlation analysis algorithm.4.Image recognition based on multi-kernel sparsity preserving canonical correlation analysis.Although KSPCCA algorithm overcomes the shortcomings of the CCA algorithm without using sample class information and the problem of extracting nonlinear features.However,there is still a problem that the change of kernel parameters has a significant influence on the recognition rate.To solve this problem,we propose the multi-kernel sparsity preserving canonical correlation analysis.The algorithm uses different kernel functions SPCCA algorithm of data samples for learning and integration,get the discriminative features to do data learning and integration,and get the discriminative features for classification and recognition.The experimental results on JAFFE,Yale,PICS,ORL face database and UCI handwritten digit database show that this method has good recognition rate.Compared with other related algorithms,the results show that the proposed algorithm has certain advantages in terms of extracting nonlinear feature.