Robust deflated canonical correlation analysis via feature factoring for multi-view image classification

Canonical Correlation Analysis (CCA) and its kernel versions (KCCA) are well-known techniques adopted in feature representation and classification for images. However, their performances are significantly affected when the images are noisy and in multiple views. In this paper, the method of robust deflated canonical correlation analysis via feature factoring for multi-view image classification is proposed. In this method, a feature factoring matrix is introduced to measure proximities between each feature vector in the dimension and projection vector, through this we evaluate the contribution of each feature to the whole feature space. Therefore, we can assign specific weights to different features accordingly to suppress the noisy data. As images are captured in multi-view usually, we also propose a deflated CCA method to build multiple factoring matrices with respect to multiple projection vectors. In this way, we weigh the degree of importance of each feature in each projection respectively to get a better feature representation for multi-view images. Experimental results on several datasets such as ORL, COIL and USPS, demonstrate that our method can improve classification performance compared to other state-of-the-art CCA methods.

Automated summary

The paper proposes a method of robust deflated canonical correlation analysis via feature factoring for multi-view image classification, which introduces a feature factoring matrix to evaluate the contribution of each feature to the whole feature space and assign specific weights to different features to suppress noisy data. Multiple factoring matrices are built with respect to multiple projection vectors to weigh the degree of importance of each feature in each projection for better feature representation in multi-view images.