New face recognition algorithm based on nonlinear dimensionality reduction

Abstract

A new supervised kernel subspace method for face recognition was proposed. It makes a better use of samples' label information by introducing supervision in the construction of local neighborhood so as to improve the recognition efficiency; and a positive regularization was added in the gaining of the optimal reconstruction weight matrix to make it insensitive to the noise. In experimental phase, the ATT and Yale face databases were used to test this algorithm by using the Nearest Neighborhood (NN) algorithm to construct the classifiers. Experimental results show that the proposed method is effective, robust and is superior to non-supervised KNPP and the traditional classic methods.