IR and visible-light face recognition using canonical correlation analysis

This paper proposes a novel multispectral feature extraction method according to the idea of canonical correlation analysis (CCA). Instead of extracting two groups of features with the same pattern (modality) as usual, the work explores another type of application of CCA that for extracting most correlated features from different face modalities to form effective discriminant vectors for recognition. Our goal is to search the complementary information in visible-light and infrared (IR) face imagery that are insensitive to the variation in expression and in illumination. Experimental results on Notre Dame face database show that the proposed CCA-based multispectral algorithm outperforms previous methods using visible-light imagery.