Prior Information Based Canonical Correlation Analysis And Its Applications

In pattern recognition, the same objects are often described by different views. Generally,different views can reflect different statistical information of the same objects, which is complementary each other. Canonical correlation analysis(CCA) is a classical but still powerful tool for analyzing multiple view data, which is mainly used to reveal the correlation between two views of samples. The goal of CCA is to find two groups of projection directions,which can make two groups of canonical projections maximally correlational. After feature extraction, canonical correlation based methods can not only keep the effective discriminative information among features, but also eliminate redundant information between them to some extent, thus reducing the complexity of recognition algorithms. Essentially, CCA is an unsupervised subspace learning algorithm, which does not utilize the supervised information of samples. Therefore, this paper introduces the prior information to CCA, and deeply studies the theory of supervised CCA and its applications to face recognition, handwriting digit recognition, and object recognition. The main content and innovative results are as follows:1. A novel canonical correlation analysis algorithm is proposed. Without changing the original meaning of canonical correlation, the algorithm simultaneously considers the class information of within-view and between-view training samples, so that there is the greatest correlation between samples that have the same labels. The proposed method is applied to face and general object image recognition. The experimental results on the AT&T and Yale-B face image databases and the COIL-20 object image database show our proposed algorithm provides better recognition results on the whole than existing feature extraction methods.2. A relative strength based supervised CCA algorithm is proposed. The proposed algorithm not only makes the correlation between two-set canonical projections maximum,but also their every component capable of describing the relative strength between different samples. The extracted canonical features by our algorithm are very discriminative in classification due to the use of the strategy that the difference between inter-class samples is larger than that between intra-class samples. Many experimental results on handwritten digit,face, and object image datasets show that the proposed algorithm has better performance than existing canonical correlation feature extraction algorithms.3. A novel spectrum decomposition multi-view learning framework is proposed by defining scatter requirement. Many existing subspace learning algorithms can be regarded as its special cases. In this framework, six covariance matrices are first designed from the between-view, within-view, between-class, within-classes, and overall viewpoints. Then, these six matrices are sorted via using scatter requirement standard. At last, multiple kinds of multi-view learning algorithms based on spectrum decomposition are presented. In addition,we further give the physical meaning from the perspective of discriminant analysis and correlation analysis. A lot of expriments on handwritten digit, face, and object image datasets demonstrate the effectiveness of the proposed framework.