Enhanced Linear Discriminant Canonical Correlation Analysis for Cross-modal Fusion Recognition.

Based on discriminant canonical correlation analysis of LDA, a new method of multimodal information analysis and fusion is proposed in this paper. We process data from two perspectives, single modality and cross-modal. More specifically, firstly, LDA is utilised to obtain the best projection matrix, this way, the data in each within-modal can be as centralized as possible. Secondly, the improved DCCA is used to process the output of first step in order to maximize within-class correlation and minimize between-class correlation. The above two steps prove beneficial to obtain the feature with higher discriminating ability which is essential for the average fusion recognition accuracy improvement. We show state-of-art results or better than state-of-art on widely used USM benchmarks against all existing results include CCA, LDA, DCCA, GCCA and KCCA.