A Discriminant Two-Dimensional Canonical Correlation Analysis

Two-dimensional canonical correlation analysis (2D-CCA) is an effective method for two-view feature extraction and fusion. It is able to reduce the computational complexity while reserving local data structure of a two-dimensional signal, e.g., image. However, since 2D-CCA only reveals the structure of the input data set without discriminatory information, it cannot effectively extract and represent discriminant representations for recognition and classification. Aiming at providing a method for discriminant feature extraction and fusion, this paper proposes a discriminant 2D canonical correlation analysis (D2DCCA), utilizing descriptor of 2DCCA and scatter information of different classes to improve recognition performance. We conduct experiments on AR face database to evaluate the performance of the proposed method. Experimental results show that the proposed D2DCCA outperforms other related algorithms.