Improved canonical correlation analysis and its applications in image recognition

The objective function of traditional canonical correlation analysis (CCA) is to maximize the correlations of the two observations of the same patterns, which can be regarded as a unsupervised learning method. Based on the traditional CCA, by introducing the class overall correlation matrix to improve the objective function, a new improved canonical correlation analysis (ICCA) method of multi-mode feature extraction is proposed, which is a supervised learning algorithm. The theory of ICCA can be explained as follows: If a pattern has a pair of observations (For any pattern space, there has two observation spaces), ICCA can find a relevant subspace of the two observation spaces, in which the mappings of the pair observations have maximum correlation, and in the relevant subspace, the mappings can have more discriminate semantic information for pattern recognition. The process of ICCA can be divided into three steps: extract two groups of feature vectors with the same pattern; establish the objective function of maximizing the class overall correlations of the all classes; extract their improved canonical correlation features to form effective discriminate vectors for recognition. Our proposed algorithm ICCA is validated by the experiments on Yale face database. Be compare with other methods, the recognition rate of our method is far higher than that of PCA algorithm only adopting single-mode features and the traditional CCA multi-mode feature fusion algorithm. © 2010 Binary Information Press.