A Simple and Effective Discriminant Feature Construction Method for Image Recognition

Raw data from real-world applications are usually high-dimensional data with noise and redundancy information. How to obtain low-dimensional features of raw data is crucial to pattern recognition. In this paper, we propose a simple and effective discriminant feature construction (DFC) method, which exploits class labels and nonlinear similarity information to directly construct discriminant similarity features of training samples. The features keep the discriminating power of class labels and similarity information as much as possible.