TWO-VIEW MEDIAN CORRELATION ANALYSIS

Abstract

Canonical correlation analysis based on supervised information is able to learn discriminant correlation features from two-view data, which plays an important role in pattern recognition and machine learning. However, such methods mainly employ class means that are sensitive to outlier data. To solve the issue, we propose a robust two-view feature learning method, called two-view median correlation analysis. In the method, a discriminant median scatter of each view is constructed in order to enhance the robustness of outlier data, and we learn correlation features with well class separability by further constraining the discriminant median scatters on the basis of maximum between-view correlation. Promising experiment results have demonstrated the effectiveness of our method.