Distance preserving multi-modal correlation feature learning with applications to image recognition

Abstract

Canonical Correlation Analysis (CCA) is a famous feature learning method that learns a pair of correlation projection directions by maximizing correlation coefficients between different modalities. However, CCA ignores nonlinear discriminative geometry structure information hidden in multi-modal data, and the ignored information is crucial to the recognition performance of feature learning. Aimed at this issue, we propose a novel multi-modal correlation feature learning method, i.e. distance preserving supervised correlation analysis (DPSCA). In this method, we construct discriminative distance scatters for capturing distance-based discriminative geometry structures. By maximizing the between-modal correlation and simultaneously constraining the discriminative distance scatters, correlation features learned by DPSCA can effectively preserve the structures, which will be beneficial to enhance the discriminative power of the learned correlation features. Excellent experimental results on real image datasets prove the superiority of the method.