Two-phase linear reconstruction measure-based classification for face recognition

In this article we propose several two-phase representation-based classification (RBC) methods that are inspired by the idea of the two-phase test sample sparse representation (TPTSR) method with L₂-norm. We first introduce two simple extensions of TPTSR using L₁-norm alone and the combination of L₁-norm and L₂-norm, respectively. We then propose two-phase linear reconstruction measure-based classification (TPLRMC) by adopting the linear reconstruction measure (LRM). Decomposing each feature sample as a weighted linear combination of the other feature samples, TPLRMC can measure the similarities between any pairs of feature samples. The linear reconstruction coefficients can capture the feature's neighborhood structure that is hidden in data. Thus, these coefficients with L_p-norm regularization can be used as good similarity measures between samples and the test ones in classifier design of TPLRMC to enhance discriminative capability. In regard to the classification procedure, TPLRMC first coarsely searches K nearest neighbors for a given query sample with LRM, then finely represents the query sample as a linear combination of the chosen K nearest neighbors, and finally uses LRM to perform classification. The experimental results on six face recognition databases and two object recognition databases demonstrate that the proposed methods outperform the competitors used in the experiments.