A Novel Local Sensitive Frontier Analysis for Feature Extraction

In this paper, an efficient feature extraction method, named local sensitive frontier analysis (LSFA), is proposed. LSFA tries to find instances near the crossing of the multi-manifold, which are sensitive to classification, to form the frontier automatically. For each frontier pairwise, those belonging to the same class are applied to construct the sensitive within-class scatter; otherwise, they are applied to form the sensitive between-class scatter. In order to improve the discriminant ability of the instances in low dimensional subspace, a set of optimal projection vectors has been explored to maximize the trace of the sensitive within-class scatter and simultaneously, to minimize the trace of the sensitive between-class scatter. Moreover, with comparisons to some unsupervised methods, such as Unsupervised Discriminant Projection (UDP), as well as some other supervised feature extraction techniques, for example Linear Discriminant Analysis (LDA) and Locality Sensitive Discriminant Analysis (LSDA), the proposed method obtains better performance, which has been validated by the results of the experiments on YALE face database.