Face Recognition Based on Improved PCA Reconstruction

A face recognition method based on improved principal components analysis (PCA) reconstruction is proposed. Firstly, PCA algorithm was performed on training samples of each pattern class to calculate the optimal projection transformation matrices. A point that should be mentioned was that we used median vector rather than mean vector in total scatter matrix. The feature vectors of testing sample could be obtained by projecting it on the optimal projection transformation matrices. After that, reconstruction images phase was conducted to get the reconstruction image. Using the same procedure, the reconstruction image of testing image corresponding to each pattern class could be obtained. Finally, the error between reconstruction images and testing sample were calculated, respectively. The testing sample was belonging to the pattern class whose corresponding error was minimal. Experiments on Yale and ORL show that this approach works much better than traditional PCA.