Sparse Fuzzy Two-Dimensional Discriminant Local Preserving Projection (SF2DDLPP) for Robust Image Feature Extraction

Abstract Recently, image feature extraction algorithms based on 2D discriminant local preserving projection (2DDLPP) algorithms have been successfully applied in many fields. The 2DDLPP can maintain the discrimination information of the local intrinsic manifold structure using two-dimensional image representation data. However, the 2DDLPP algorithm encounters the problem of the sensitivity of overlapping points (outliers) and requires high computational cost in real-world applications. In order to resolve the problems mentioned above, we introduce a new elastic feature extraction algorithm called the sparse fuzzy 2D discriminant local preserving projection (SF2DDLPP). First, the membership matrix is calculated using the fuzzy k-nearest neighbours (FKNN), which is applied to the intraclass weighted matrix and the interclass weighted matrix. Second, two theorems are developed to directly solve the generalized eigenfunctions. Finally, the optimal sparse fuzzy 2D discriminant projection matrix is regressed using the elastic net regression. The experiments show the effectiveness and stability of this algorithm on several face (ORL, Yale, AR and Yale B), USPS and palm print datasets.