Research On Feature Extraction And Image Recognition Based On Correlation Projection Analysis

Feature extraction is one of the most basic problems of pattern recognition. For image recognition tasks, extracting the effective image feature is the key step. The correlation projection analysis proposed in this paper, including Canonical Correlation Analysis, Generalized Canonical Correlation Analysis and Partial Least Squares Analysis, is a feature extraction method based on two groups of features or two data matrices. Its basic idea is: first, establish correlation criterion function between two groups of features; then according to the criterion function, solve two groups of correlation projection vectors, and extract the correlation feature of each data; finally, through different feature fusion strategies, obtain the combined correlation feature, which is used in image classification. This paper deeply discusses the theory and algorithms of these three correlation projection analysis, and the proposed algorithms are successfully applied in face recognition and handwritten character recognition.Based on the idea of feature fusion, the framework of Canonical Correlation Analysis (CCA) used in combine feature extraction and pattern recognition are discussed. The proposed two kinds of feature fusion strategies make a new way of feature extraction and classification of pattern. In this paper, not only the theory and methods of CCA applied in image recognition are provided, what is more important is that the application of CCA is theoretically extended. It solves the problem of canonical projection vectors when two total covariance matrices formed by two groups of feature vectors are singular, such that it fits for the case of high-dimensional space and small sample size. Furthermore this paper presents the relationship of CCA and Fisher linear discriminant analysis, and under the framework of canonical projection vectors, a new algorithm of solving uncorrelation optimal discriminant vectors is proposed. The proposed algorithms are validated on the handwritten Arabic numerals database and face database.From the angle of favoring pattern classification, this paper proposes and discussed the theory and algorithms of Generalized Canonical Correlation Analysis (GCCA) in detail. First, an improved correlation criterion function is proposed. According to this criterion, the algorithm theory of solving the generalized canonical projective vectors(GCPV) is respectively built, which is under two kinds of constraints. Furthermore, when the within-class scatter matrices of two groups of the original sample are singular, two strategies of solving GCPV are discussed, namely PCA+GCCA strategy and Perturbation strategy. To solve GCPV using Perturbation strategy, a high efficient algorithm is proposed aiming to computational efficiency of high-dimensional space and small sample size problems (such as image recognition). A new feature fusion strategy 3 (FFS3) is proposed in this paper. It is introduced the concept of correlation feature matrix, and two kinds of classification methods based on correlation feature matrix are proposed. Experimental results on the handwritten Arabic numerals database and face database validate the proposed algorithms. The theoretical analysis and experimental results both show that based on the improved criterion function, it not only has an obvious physical meaning, but also according to the classification capability, the combine feature extracted by GCCA is...