Application to Three-Dimensional Canonical Correlation Analysis for Feature Fusion in Image Recognition

This paper presents a three-dimensional canonical correlation analysis (TCCA) method, and applies it to feature fusion for image recognition. It is an extension of traditional canonical correlation analysis (CCA) and two-dimensional canonical correlation analysis (2DCCA). Considering two views of a three-dimensional data, the TCCA can directly find the relations between them without reshaping the data into matrices or vectors, We stress that TCCA dramatically reduce the computational complexity, compared to the CCA and 2DCCA. To evaluate the algorithm, we are using Gabor wavelet to generate the three-dimensional data, and fusing them at the feature level by TCCA. Some experiments on ORL database and JAFEE database and compared with other methods, the results show that the TCCA not only the computing complexity is lower, the recognition performance is better but also suitable for data fusion.