Label-wise Orthogonal Canonical Correlation Analysis and Its Application to Image Recognition

This paper proposes the label-wise orthogonal canonical correlation analysis (LOCCA), which constrains the label-based relationships and orthogonalizes correlation projection directions. In the method, the discriminative structures constrained by class labels are effectively preserved, and the correlation projection directions from LOCCA reduce the information redundancy by orthogonality criterion as much as possible. Encouraging experimental results on two real-world image datasets reflect that our method is an effective and robust image recognition method.