Recognition based on wavelet reconstruction face

Face recognition algorithms have to deal with significant amounts of illumination and expression variations between gallery and probe images. This paper analyzes the facial image of multi-level wavelet decomposition features, points out the facts that illumination variations have the greatest impact on the low-frequency decomposition approximation coefficients, followed by expression and individual changes. And a novel face reconstruction method is proposed. The method firstly decomposes the face image by multi-level wavelet, and projects the low-frequency approximation coefficients onto the subspace, which is made from the normal face samples by PCA. Then it selects the illumination unrelated coefficients to rebuild the low-frequency approximation coefficients to replace the original ones. After wavelet construction, we can get the illumination unrelated face. The followed experiments which based on the classic eigenface algorithm show it can not only decrease the illumination and expression impacts on face image and improve the recognition rate greatly (28.9%), but also make the results robust to the change of eigenspace dimension.