Kernel Generalized Canonical Correlation and a New Feature Fusion Strategy

Abstract

When processing face images, we usually use the data analysis tools to find the underlying relations in data. As an extraction method based on two groups of features, Canonical Correlation Analysis (CCA) can better reveal the inner relations between two variates. Recently, CCA has been widely used in multiple feature fusion and extraction and it has gained a series of achievements. The basic idea of CCA is: firstly, establish correlation criterion function between two groups of features; then extract the correlation feature of each data; finally, obtain the combined correlation feature, which is used in image classification. In practice, we can also add the category information of training samples by improving the discriminant criterion function. Furthermore, by exploiting corresponding kernel function, the original data can also be projected into higher dimensional spaces to reduce the gaps between the heterogeneous spaces. In this paper, we focus on the Generalized CCA(GCCA) and Kernel CCA(KCCA) and then introduce the idea of kernel GCCA(KGCCA). We solve the problem by Lagrangian multiplier method and we also propose a new feature fusion strategy (FFS). Finally, experiments on Yale and ORL face datasets demonstrate the performance of our proposed method KGCCA and the new FFS.