Feature Extraction via Balanced Average Neighborhood Margin Maximization

摘要

Average Neighborhood Margin Maximization (ANMM) is an effective method for feature extraction, especially for addressing the Small Sample Size (SSS) problem. For each specific training sample, ANMM enlarges the margin between itself and its neighbors which are not in its class (heterogeneous neighbors), meanwhile keeps this training sample and its neighbors which belong to the same class (homogeneous neighbor) as close as possible. However, these two requirements are sometimes conflicting in practice. For the purpose of balancing these conflicting requirements and discovering the side information for both the homogeneous neighborhood and the heterogeneous neighborhood, we propose a new type of ANMM in this paper, called Balance ANMM (BANMM). The proposed algorithm not only can enhance the discriminative ability of ANMM, but also can preserve the local structure of training data. Experiments conducted on three well-known face databases i.e. Yale, YaleB and CMU PIE demonstrate the proposed algorithm outperforms ANMM in all three data sets.