Locality-regularized linear regression discriminant analysis for feature extraction

Pu Huang, Tao Li, Zhenqiu Shu, Guangwei Gao, Geng Yang, Chengshan Qian

Abstract: Locality-regularized linear regression classification (LLRC) is an effective classifier that shows great potential for face recognition. However, the original feature space cannot guarantee the classification efficiency of LLRC. To alleviate this problem, we propose a novel dimensionality reduction method called locality-regularized linear regression discriminant analysis (LLRDA) for feature extraction. The proposed LLRDA is developed according to the decision rule of LLRC and seeks to generate a subspace that is discriminant for LLRC. Specifically, the intra-class and inter-class local reconstruction scatters are first defined to characterize the compactness and separability of samples, respectively. Then, the objective function for LLRDA is derived by maximizing the inter-class local reconstruction scatter and simultaneously minimizing the intra-class local reconstruction scatter. Extensive experimental results on CMU PIE, ORL, FERET, and Yale-B face databases validate the effectiveness of our proposed method.