Local Two-Dimensional Canonical Correlation Analysis

Recently, two-dimensional canonical correlation analysis (2DCCA) has been proposed for image analysis. 2DCCA seeks linear correlation based on images directly. It fails to identify nonlinear correlation between two sets of images. In this letter, we present a new manifold learning method called local 2DCCA (L2DCCA) to identify the local correlation. Different from 2DCCA in which images are globally equally treated, L2DCCA weights images differently according to their closeness. That is, the correlation is measured locally, which makes L2DCCA more accurate in finding correlative information. Computationally, L2DCCA is formulated as solving generalized eigenvalue equations tuned by Laplacian matrices. Like 2DCCA, the implementation of L2DCCA is straightforward. Experiments on FERET and UMIST face databases show the effectiveness of the proposed method.