Analysis on two fisher methods and a synthesized discriminant projection

This paper proposes an extension of marginal fisher analysis (EMFA) for dimensionality reduction and analyzes some properties of both EMFA and linear discriminant analysis (LDA), and finally suggests a synthesized discriminant projection (SDP). SDP takes both global class relationship and local geometry structure into account, which maximizes the distance between marginal points and the distance between different class-centers, thus enhances classification.