Metric learning-based kernel transformer with triplets and label constraints for feature fusion

Highlights：

• First, we propose a kernel transformer to learn the complementary properties of different features for feature fusion. Based on the Mahalanobis distance, for feature concatenation in kernel space, we prove that the fusion can be achieved in the data space using each feature.

• Second, kernel metric learning with triplets and label constraints is proposed, which leads to a better performance compared with the method that only triplets constraints are used. Based on the theory of extreme learning machine, label constraints are also embedded into our model.

• Third, we built a complete optimization objective function. Based on the alternating direction method of multipliers solver and the Karush-Kuhn-Tucker theorem, the proposed optimization problem is solved with rigorous theoretical analysis.