A Subspace Method Based on Data Generation Model with Class Information

摘要

Subspace methods have been used widely for reduction capacity of memory or complexity of system and increasing classification performances in pattern recognition and signal processing. We propose a new subspace method based on a data generation model with intra-class factor and extra-class factor. The extra-class factor is associated with the distribution of classes and is important for discriminating classes. The intra-class factor is associated with the distribution within a class, and is required to be diminished for obtaining high class-separability. In the proposed method, we first estimate the intra-class factors and reduce them from the original data. We then extract the extra-class factors by PCA. For verification of proposed method, we conducted computational experiments on real facial data, and show that it gives better performance than conventional methods.